[Improvement] max_batches support to training log and tqdm progress bar.

[hakuryuu96]

Issue description

As @BloodAxe described, if training_hyperparams.max_train/valid_batches are redefined in CLI, tqdm and training log do not take changes into account and continue showing the full length of the dataloader. E.g. when user executes something like this

train_params = {
    "max_epochs": 300,
    "phase_callbacks": phase_callbacks,
    "initial_lr": lr,
    "loss": loss_fn,
    "optimizer": optimizer,
    "train_metrics_list": [Accuracy(), Top5()],
    "valid_metrics_list": [Accuracy(), Top5()],
    "metric_to_watch": "Accuracy",
    "greater_metric_to_watch_is_better": True,
    "lr_scheduler_step_type": "epoch",
    "max_train_batches": 24,
    "max_valid_batches": 24,
}

trainer.train(model=net, training_params=train_params, train_loader=train_loader, valid_loader=valid_loader)

the resulting logs and progress bar are the following:

[2023-10-19 13:30:41] INFO - sg_trainer_utils.py - TRAINING PARAMETERS:
    - Mode:                         Single GPU
    - Number of GPUs:               1          (2 available on the machine)
    - Full dataset size:            50000      (len(train_set))
    - Batch size per GPU:           16         (batch_size)
    - Batch Accumulate:             1          (batch_accumulate)
    - Total batch size:             16         (num_gpus * batch_size)
    - Effective Batch size:         16         (num_gpus * batch_size * batch_accumulate)
    - Iterations per epoch:         3125         (len(train_loader))
    - Gradient updates per epoch:   3125         (len(train_loader) / batch_accumulate)

[2023-10-19 13:30:41] WARNING - sg_trainer_utils.py - max_train_batch is set to 24. This limits the number of iterations per epoch and gradient updates per epoch.
[2023-10-19 13:30:41] INFO - sg_trainer.py - Started training for 300 epochs (0/299)

Train epoch 0: 1%|█                        | 24/3125 [00:02<00:00,  9.22it/s, Accuracy=0.0729, CrossEntropyLoss=2.53, Top5=0.508, gpu_mem=0.231]
Validating: 1%|█                      | 24/3125 [00:00<00:00, 126.12it/s]
2023-10-19 13:30:44] INFO - base_sg_logger.py - Checkpoint saved in /home/phil/deci/super-gradients/checkpoints/Cifar10_external_objects_example/RUN_20231019_133041_486426/ckpt_best.pth
[2023-10-19 13:30:44] INFO - sg_trainer.py - Best checkpoint overriden: validation Accuracy: 0.1002604141831398
===========================================================
SUMMARY OF EPOCH 0
├── Train
│   ├── Crossentropyloss = 2.5337
│   ├── Accuracy = 0.0729
│   └── Top5 = 0.5078
└── Validation
    ├── Crossentropyloss = 2.3486
    ├── Accuracy = 0.1003
    └── Top5 = 0.4596

===========================================================

PR description

This PR addresses the issue above and proposes some improvements. Briefly:

1. Logs show the actual number of used elements in dataloader
   1.5. Additionally, logs are warning user that the max_batches parameter was set.
2. Progress bar shows the actual number of steps it will take to finish the epoch

E.g. if the max_train/valid_batches parameter is specified:

[2023-10-19 13:30:41] INFO - sg_trainer_utils.py - TRAINING PARAMETERS:
    - Mode:                         Single GPU
    - Number of GPUs:               1          (2 available on the machine)
    - Full dataset size:            50000      (len(train_set))
    - Batch size per GPU:           16         (batch_size)
    - Batch Accumulate:             1          (batch_accumulate)
    - Total batch size:             16         (num_gpus * batch_size)
    - Effective Batch size:         16         (num_gpus * batch_size * batch_accumulate)
    - Iterations per epoch:         24         (len(train_loader) OR max_train_batches)
    - Gradient updates per epoch:   24         (len(train_loader) OR max_train_batches / batch_accumulate)

[2023-10-19 13:30:41] WARNING - sg_trainer_utils.py - max_train_batch is set to 24. This limits the number of iterations per epoch and gradient updates per epoch.
[2023-10-19 13:30:41] INFO - sg_trainer.py - Started training for 300 epochs (0/299)

Train epoch 0: 100%|██████████| 24/24 [00:02<00:00,  9.22it/s, Accuracy=0.0729, CrossEntropyLoss=2.53, Top5=0.508, gpu_mem=0.231]
Validating: 100%|██████████| 24/24 [00:00<00:00, 126.12it/s]
[2023-10-19 13:30:44] INFO - base_sg_logger.py - Checkpoint saved in /home/phil/deci/super-gradients/checkpoints/Cifar10_external_objects_example/RUN_20231019_133041_486426/ckpt_best.pth
[2023-10-19 13:30:44] INFO - sg_trainer.py - Best checkpoint overriden: validation Accuracy: 0.1002604141831398

If not, the logs behave similarly to previous versions.

Some ideas

IMO it should be cool to consider logging the whole set of training parameters before the run. For me as a user, it would be nice to double-check all the settings I've made somewhere in the project (e.g. if I use hydra and take SG Trainer class to my pipeline) and to be sure things go smoothly :)
For example, the user should see the following info:

• dataset (number of classes, class names, etc)
• dataloader (batch_size, num_workers, etc)
• model (model name, number of trainable parameters, etc)
• optimization info (optimizer, lr, wd, losses, etc)
• training info (number of epochs, number of gradient updates, EMA, SyncBN usage, etc)

[BloodAxe]

LGTM

[Louis-Dupont]

LGTM