feat: exclude refit sensitive ops from TRT compilation

[peri044]

Description

1. Fix refit errors on certain ops by forcing them to fallback on pytorch. The global default for make_refittable is set to False even after this addition.
2. Adds a capability validator for cumsum and embedding bag to let them fallback to pytorch incase user builds refittable engines
3. Expose user compilation settings to ConverterRegistry
4. Added a testcase to verify fallback
5. The converter testcases for cumsum and embedding bag are intentionally marked as make_refittable=False. This is because in the future, we want make_refitable=True by default and hence these tests fail. So, explicitly marking these now will avoid those errors.

Fixes https://github.com/pytorch/TensorRT/actions/runs/10637055400/job/29490744653?pr=3131

Type of change

Please delete options that are not relevant and/or add your own.

• Bug fix (non-breaking change which fixes an issue)
• New feature (non-breaking change which adds functionality)
• Breaking change (fix or feature that would cause existing functionality to not work as expected)
• This change requires a documentation update

Checklist:

• My code follows the style guidelines of this project (You can use the linters)
• I have performed a self-review of my own code
• I have commented my code, particularly in hard-to-understand areas and hacks
• I have made corresponding changes to the documentation
• I have added tests to verify my fix or my feature
• New and existing unit tests pass locally with my changes
• I have added the relevant labels to my PR in so that relevant reviewers are notified

[narendasan]

I think its better to do this through validators than maintaining another list and modifying user settings

[peri044]

1. Currently, the capability validator doesn't have access to user settings. So, I think making the compilation settings in _compiler.py (compile call) a global variable and passing it here would work. https://github.com/pytorch/TensorRT/blob/main/py/torch_tensorrt/dynamo/conversion/_ConverterRegistry.py#L439

2. The other alternative if we don't want to modify user settings is

CONVERTERS.set_disallowed_targets(settings.REFIT_SENSITIVE_OPS)

What would be preferred ?

[narendasan]

Dont see a drawback to let validators see user settings, could be relevant later like if a certain converter only works for one data type

[narendasan]

So it would involve updating the converter registry to take the settings as an arg in addition to the node, and some adjustment in the decorator. But overall I think its the best design since we shouldnt be spreading this info in a bunch of global lists, a converter should be able to tell you what it can and cant do on its own

Dont think we need to make settings global we could make a reference a member of the registry like we do the dynamic shape setting

TensorRT/py/torch_tensorrt/dynamo/conversion/_ConverterRegistry.py

Line 300 in 4d2a04a

assume_dynamic_shape_support: bool = False,

and then it gets injected into the call to the validator

[narendasan]

Just an additional arg here:

TensorRT/py/torch_tensorrt/dynamo/conversion/_ConverterRegistry.py

Line 439 in 4d2a04a

if candidate.capability_validator(node) and (

[peri044]

Got it. modified it now.

[narendasan]

LGTM, make sure to mark needs cherrypick and ping @lanluo-nvidia