refactor(pytorch): clarify scheduler and input-maker control flow

[grimoire]

Description
This PR is a readability-focused refactor of the PyTorch scheduler/input-maker path. It keeps scheduling behavior UNCHANGED while making long-context chunk planning, forward input selection, prefill admission, and BlockTrie ownership easier to reason about.

What Changed

• Added characterization tests for long-context chunk reservation contracts.
• Shared long-context chunk planning between scheduler and input maker.
• Extracted _ForwardInputsTask from InputsMakerAsync._make_forward_inputs() to separate selection policy from payload construction.
• Split scheduler prefill reorder/admission flow into _PrefillReorderer and _PrefillAdmissionAttempt.
• Clarified BlockTrie ownership, SSM checkpoint lifecycle, eviction, rollback, and leaf bookkeeping with docstrings/comments.
• Added warning-only diagnostics for impossible internal states.

Non-Goals

• No new scheduling feature.
• No intended behavior change.