fix(tests): compare baddbmm lowering against TensorRT in fp32, not TF32

[shoumikhin]

Problem

dynamo-lowering-rtx is red on main:

FAILED lowering/test_decompositions.py::TestLowering::test_lowering_baddbmm
AssertionError: 0.0019378662109375 != 0 within 4 places
: baddbmm TRT outputs don't match with the original model.

The same test passes on standard TensorRT, so it went in green and only turned red once the RTX lane ran on main.

Cause

The test compares a TensorRT engine against an eager fp32 reference to DECIMALS_OF_AGREEMENT, but builds the engine with TF32 left at its default, which is on. TF32 rounds both matmul operands to 10 mantissa bits before multiplying. A passing run therefore only means TensorRT happened to pick a kernel that was not TF32; it is not something the test asks for. TensorRT-RTX picks a TF32 kernel and the check fails.

The decomposition is not at fault. Over 500 random draws of the shapes the test uses, bias + bmm(batch1, batch2) in fp32 matches aten.baddbmm bit for bit (max difference 0.0). Repeating the same math with the two operands rounded to 10 mantissa bits drifts by up to 0.0054, which brackets the 0.0019 seen in CI.

Fix

Pass disable_tf32=True so both sides run the same arithmetic. The tolerance is unchanged and nothing is skipped. This matches how other numeric tests in the repo handle it, for example tests/py/dynamo/conversion/test_index_put_aten.py and tests/py/dynamo/models/test_models.py.

Testing

test_lowering_baddbmm passes locally on standard TensorRT with the change, and the debug log confirms disable_tf32=True reaches the compilation settings through the torch_compile path. The RTX lane needs ci: full or an approving review to run here.