[PyTorch] Refactor GroupedLinear eager forward and backward

[pggPL]

Description

Extract GroupedLinear forward/backward computation and saved-state handling from the autograd wrapper. Both split-quantize and fused grouped-tensor execution expose shared helpers that the compile integration in #32 can reuse.

This is the first of two stacked PRs; it contains the eager refactor.

Type of change

• Code refactoring

Changes

• Introduce explicit forward/backward argument containers and thin autograd wrappers for both execution paths. Share backward state cleanup and MCore main-grad handling; omit unused argument fields.
• Move computation and context setup into shared functions, including fused grouped GEMM.
• Preserve eager saved-tensor layout, packed gradients, grouped parameters, save-original-input, delayed wgrad and main-grad accumulation.

Validation

RTX Ada workstation, current source:

• test_grouped_linear.py: 718 passed, 558 skipped.
• 12 comparisons against main across BF16/FP8, save-original-input and gradient settings: bit-exact outputs/gradients and identical saved-tensor shapes, strides, offsets and storage-alias groups.
• 72 BF16 fused comparisons against main across grouped weight/bias flags, save-original-input, gradient settings and normal/main-grad/delayed wgrad: bit-exact values and saved metadata. Grouped GEMM was replaced with matmuls for this temporary probe.
• Changed-file pylint, Black and diff checks pass.

Native fused GEMM requires Hopper/Blackwell and was not exercised on RTX Ada. Full Python lint reports the existing import-order diagnostic in unchanged attention/fused_mla_q_uproj.py.

Checklist

• I have read and followed the contributing guidelines
• Existing eager tests pass on the available GPU
• Changes introduce no lint diagnostics
• Compile integration is isolated in the follow-up PR