skip dummy inference and run_shape_analysis

[lanluo-nvidia]

Description

There is two changes introduced in this PR:

1. during the compile stage:
   skipped dummy inference and use graph inspection instead to get the output_node.meta['val']

2. during the save stage:
   skipped run_shape_analysis and use graph inspection instead to get the output_node.meta['val']

Fixes # (issue)

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

[peri044]

Here's what I think could be a simpler way of doing this

1. We probably don't have to store output_shapes in TorchTensorRTModule class. Once the compilation is finished, verify if the nodes of the TRT graph modules have metadata in them (if not we can update it by node.meta["val"] = original metadata)
   Reference:

   TensorRT/py/torch_tensorrt/dynamo/_compiler.py

   Lines 496 to 499 in 3eb48d7

   	for name, trt_module in trt_modules.items():
   	setattr(partitioned_module, name, trt_module)
   	if settings.lazy_engine_init:
   	getattr(partitioned_module, name).setup_engine()

Here the node corresponding to _run_on_acc0 can be queried as

trt_module_node = [node for node in gm.graph.nodes if node.name == "_run_on_acc0"]
trt_module_node.meta["val"] - This should already have fake tensors which need to be used in the exporter. 

2. exporter
   We have the TRT module node here :

   TensorRT/py/torch_tensorrt/dynamo/_exporter.py

   Line 364 in 3eb48d7

   trt_module_node = [node for node in gm.graph.nodes if node.name == name]

   
   We could directly set (ensuring trt_module_node.meta["val"] always exists)

trt_node.meta["val"] = trt_module_node.meta["val"]

3. infer_module_types
   We can replace the dummy inference with graph inspection by reading output metadata. The output of this function could be a list of FakeTensors and we can extract the dtypes from this to pass it to TRTInterpreter.

Replacing the dummy inference will also need changes to our converter test suite.

[peri044]

Besides checking val field in node.meta, also check tensor_meta if val is not present. Apparently, torch.compile stores metadata in tensor_meta while export uses val fields in node.meta

[peri044]

Added minor comments. Mostly LGTM

[peri044]

LGTM