[Refactor] Extract GQA prefill kernel selection logic

[superAngGao]

Parent: #1610

Scope

Extract the first piece of GQA kernel selection policy while keeping the intended layering and manifest visibility explicit.

Preferred layering:

OP class -> local selector/helper -> self.kernel_map[key] -> concrete Kernel class -> TileLang implementation/factory/callable

Required design constraints

Do this:

Op -> selector/helper -> concrete Kernel -> lower-level TL implementation/factory/callable

Do not do this:

logical Kernel -> existing Kernel -> TL implementation

The OP-to-kernel relationship should remain visible through default_kernel_map and manifest source.kernel_map where applicable. Shared dispatch logic can live in a local pure helper near the GQA OP code, such as _select_gqa_prefill_kernel_key(...); it does not need to become a concrete Kernel subclass or a new module in the first patch.

Manifest constraint

This issue must not add new public OPs or change tileops/manifest/attention.yaml.

Current manifest entries have fixed ordered signature.inputs. Later OP aggregation should keep that constraint: each aggregated OP must define one explicit, fixed input list in both code and manifest. No optional tensor inputs and no dynamic arity.

Non-goals

• Do not add new public OPs.
• Do not rename existing public OPs.
• Do not change tileops/manifest/attention.yaml.
• Do not add logical Kernel subclasses.
• Do not remove existing concrete kernel classes.
• Do not hide all concrete kernels behind a single mega logical kernel that weakens manifest/kernel_map visibility.
• Do not split the large GQA files yet unless the selector extraction naturally requires a small local helper.

Expected behavior

• Dense/no-cache prefill paths can share selection logic without changing the public OP contract.
• Contiguous KV-cache prefill paths can share selection logic without changing the public OP contract.
• Paged KV-cache prefill paths can share selection logic without changing the public OP contract.
• RoPE / append-style paths remain OP-level orchestration of multiple concrete kernels.
• Dispatch tests should verify selector behavior without compiling GPU kernels.

Validation

• Unit tests for selector behavior using concrete kernel-map keys or lightweight fake maps.
• Import checks for the touched GQA OP path.
• Focused GQA smoke tests where practical.
• No manifest validation changes are expected because this issue does not add public OPs.