[QuantizationModifier] take ownership of add_observers_, unit test fixes

[bfineran]

this PR addresses three issues found in testing of the new QuantizationModifier:

1. addresses an break in recipe template tests
2. uses FakeQuantizeBase for inheritance comparisons where possible (introduced as default in later versions of torch)
3. torch 1.9 tests revealed that torch quantization will inject its own activation post process hooks into certain activations without any way to disable. this PR adds its own implementation of add observers that does not include this override. (fixes issues with quantization of Sigmoid and other activations)
4. previously, the new modifier still deferred to torch for quantization of output activations - this was a known hole as we try to quantize "everything" while torch mostly keeps to a white-list. as part of the activation fix in (3), we update our add observers implementation to quantize activations of all target modules as determined by QuantizationScheme propagation
5. fix for compatibility with the FloatFunctional module used for resnet quantization

test_plan:

• tests updated to ensure that output activation quantization is tested for existence and adherence to the scheme

[corey-nm]

torch 1.9 tests revealed that torch quantization will inject its own activation post process hooks into certain activations without any way to disable.

lol

[corey-nm]

lgtm, your version is way easier to understand than pytorch lol 🚀 💪