Bug/sg 896 add deprecation for previous breaking changes

[shaydeci]

Added deprecations, and backward compatibility to previous breaking changes.

In this PR:

• InfiniteSampler has been removed, added deprecation using equivalent DistributedSampler.

• Broken imports
  from super_gradients.training.utils.optimizers.all_optimizers import OPTIMIZERS
from super_gradients.training.models import BasicBlock, Bottleneck, HpmStruct

• DDRNetCustom: the init expected a parameter “aux_head” from “arch_params.aux_head”. In the new version of SG, this parameter is: “use_aux_heads=arch_params.use_aux_heads” - added deprecation.

• BasicDDRBackBone . In old SG version, this block didn’t expect layer3_repeats . Now SG expects it. Also, ddrnet model now assumes that “arch_params” has layer3_repeats parameter. Fixed by setting default=1.

Verified backward compatibility handled outside this PR:

Some were already handled in other PRs, this is the complete list + where they were handled:

• In ema_params, they had: exp_activation: True. Changed to decay_type: exp handled by @BloodAxe in 69a82bc
• The architecture custom_stdc name was changed to stdc_custom. Handled by @ofrimasad in 190283e
• Broken import: from super_gradients.training.models import make_divisible, handled by @BloodAxe in 6b5785d
• Had: trainer = Trainer(experiment_name=cfg.experiment_name, multi_gpu=cfg.multi_gpu). Changed to use setup_device(() instead - this is the only case where we throw an error, to avoid setting devices inside the Trainer. f50d438
• Models iinheriting from CustomizableDetector Used to call super().__init__ only with arch_params. Now its changed to accept each parameter separately:

super().__init__(
    backbone=arch_params.backbone,
    heads=arch_params.heads,
    neck=arch_params.neck,
    num_classes=arch_params.num_classes,
    bn_eps=arch_params.bn_eps,
    bn_momentum=arch_params.bn_momentum,
    inplace_act=arch_params.inplace_act
)

handled by @Louis-Dupont in a3618e8

[ofrimasad]

two minor comments

[ofrimasad]

LGTM