[PyTorch] Avoid concatenation copies for compiled split Linear parameters

[pggPL]

Description

Compiled te.Linear(parameters_split=...) can consume adjacent parameter parts
without a concatenation copy. Pass the original parts through the custom op and
construct a checked view inside the consumer, preserving separate Parameters,
state-dict names and gradients. Disjoint or noncontiguous parts use concatenation.

This PR targets upstream main and adds split-parameter compile support only to
te.Linear. ConcatInput[T] describes an ordinary operand or immutable
ParameterParts. A single concat_input helper keeps singleton tensors unchanged,
uses noop_cat in eager, and defers supported compiled operands. Original parts
cross the op boundary; fake implementations receive metadata and the real
consumer constructs the view. Linear retains its existing final configuration
validation and setup_saved_tensors(ctx) hook.

Type of change

• Non-breaking performance improvement and bug fix

Changes

• Make the early fallback check inspect original parameters without
  concatenating them; retain final validation of the prepared op arguments.
• Save original parts through autograd and split full gradients outside the
  backward op; recheck storage adjacency on each execution.
• Compute bias gradients when frozen weights skip the weight-gradient GEMM.
• Preserve dtype promotion under autocast and copy lazy negative/conjugate views.
• Keep returned or unfused split biases as traceable concatenations under compile.

Validation and remaining work

• RTX 5880 Ada, PyTorch 2.15 development build/CUDA 13.4, compatible prebuilt
  native TE extension: complete test_torch_compile.py passed (187 passed,
  46 skipped, 1 existing non-strict xfail passed). Includes 45 focused split/API
  cases covering storage and parameter replacement, CUDA Graphs, saved hooks,
  version checks, gradient targets, recipes and returned bias.
• Pre-commit, Black, full Python L0 lint (10.00/10), license and whitespace checks
  passed; tested source hashes match the commit.
• Performance work remains: disjoint storage can be concatenated repeatedly in
  backward, and earlier split-projection measurements found host overhead for
  small shapes. Ordinary-Linear CPU latency has not been benchmarked. No general
  speedup or Blackwell/distributed validation is claimed.

Checklist

• Added tests and updated the Linear parameter documentation
• Run full Python L0 lint
• Run GPU regression tests on this revision
• Complete performance validation and human review before marking ready