[zhaoyuanh-nvidia]

Description

Bring the remaining compatible review fixes from merged #20833 into feat/helix-agg, whose original phase-dependent MLA port landed in #20920. This PR is intentionally limited to four files; it does not synchronize main or rewrite the shared feature branch.

• Delegate output-slice preparation to each linear method while retaining the common range, parallelism, locality-domain, and LoRA guards.
• Keep unsupported quantized methods fail-closed and add CPU coverage for custom-method dispatch and unsupported-method rejection.
• Initialize context-head metadata in the bare MLA locality-domain test fixture and register the new CPU test.

The two low-M dispatch tests from #20833 depend on APIs not yet present on this feature branch. They belong with the separate main synchronization, not with this narrow backport. Later feature work is unchanged.

Test Coverage

• Source-extracted CPU contract probe: 27 passed, 6 native R128c4 cases deselected. This uses stub Linear construction; it is not a native TensorRT-LLM pytest or GPU run.
• Bad/fixed/restored-bad control: the same custom-method dispatch test fails on the feature base, passes with this change, and fails again when the base implementation is restored.
• Pre-commit hooks passed across the full-repository run plus the type-check retry using Git Bash. The staged-file run passed; no repository hook configuration changed. Mypy used its lightweight no-native-bindings mode (16 source files).
• Python compilation, staged-source equivalence/scope audit, and git diff --check passed.
• Native TensorRT-LLM and GPU validation remain for CI.

Integration

One DCO-signed-off commit, targeting feat/helix-agg. Merge this narrow fix before the separate main synchronization PR #21240. That follow-up preserves upstream ancestry with a merge commit rather than rewriting the shared feature branch. No automatic merge requested.