[PyTorch] Declare fusible ops as custom ops (1/3): framework + Bias

[pggPL]

Description

First step towards torch.compile(fullgraph=True) support for transformer_engine.pytorch.ops.Sequential.

The plan is to keep the pipeline-level _OperationFuserAutogradFunction and let Dynamo trace it as a
higher-order op, with each fusible operation calling its own custom op inside. That keeps the forward
and backward fusion layouts independent (so the backward-only fusions survive), keeps
OperationContext inside the traced scope, and bounds op registration to one entry per op class.

This PR does only the per-op half of that: make every basic/fused operation declarable as a
custom op, with tests. Wiring the fuser up is a separate PR, so nothing in the eager path changes yet.

Type of change

• Documentation change (change only to the documentation, either a fix or a new content)
• 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)
• Infra/Build change
• Code refactoring

Changes

• register_op_halves (dynamo/custom_op.py): registers an op's forward and backward as two
  independent two-tier custom ops and returns callables for both, without registering autograd.
  A caller wiring the halves into its own autograd.Function can then group the forward and backward
  passes differently, which is exactly what OperationFuser does. Reuses the existing
  schema/adapter/TensorSpec machinery; register_custom_op is untouched.
• Bias converted as the template: split into resolve_fwd_args (reads module config and global
  FP8 state, so it stays in the traced region where Dynamo guards those reads), pure impls that take
  everything as arguments and never touch self, and data-free fakes. op_forward / op_backward
  keep their signatures and drive the same impls, so the eager path is unchanged.
• tests/pytorch/test_ops_custom_ops.py: the fake-vs-real conformance harness every subsequent op
  will reuse, plus numerics-against-eager and fullgraph=True tests for Bias.

Note on the pass-through-gradient convention

A custom op may not return one of its own inputs. Bias's backward returns grad_output unchanged
whenever there is no grad-input quantizer, which trips that check, and cloning would cost a full-size
copy on the common path. The impl therefore returns None for that slot and the caller substitutes
grad_output. Pass-through grads are common (Identity, Reshape, ConstantScale, Quantize), so
the remaining ops will follow the same convention.

Checklist:

• I have read and followed the contributing guidelines
• The functionality is complete
• I have commented my code, particularly in hard-to-understand areas
• I have made corresponding changes to the documentation
• My changes generate no new warnings
• I have added tests that prove my fix is effective or that my feature works
• New and existing unit tests pass locally with my changes

[pggPL]

Superseded: rebased onto linear_compile_on_main and split into #29 and #30.