fix(pd): release partial request allocations on abort

[sufubao]

Summary

• follow up fix(pd): cancel pending PD requests on abort #1476 by making pending PD request cancellation exception-safe
• release every shared-memory request object and slot index acquired before cancellation
• make generation-task cancellation idempotent, remove obsolete delayed abort retries, and preserve cancellation state
• add regressions for abort-before-registration and partial slot allocation cleanup

Problem

#1476 correctly tracks and cancels generation tasks before they enter req_id_to_out_inf. However, cancellation can arrive after one or more shared-memory request indexes have been allocated while the request is still waiting for another index. The allocation then unwinds without a registered ReqStatus, so neither HttpServerManager.abort() nor the normal recycle loop can release those indexes. Repeated cancellations can permanently exhaust the finite request-slot pool.

Verification

• PYTHONPATH=$PWD python -m pytest -q unit_tests/server/httpserver (4 passed)
• Black check on all changed Python files
• flake8 on all changed Python files
• git diff --check upstream/main...HEAD

[sufubao]

Closing after rechecking the actual PD request path. Since #1439, PD master expands multi-choice requests before dispatch and sends each P/D request with n=1 and best_of=1. Therefore the partial n>1 slot-allocation scenario in this PR is not reachable from the PD abort path, and upstream #1476 is sufficient for the reported pending-abort issue. The generic allocator cleanup may still be worth evaluating separately with a real non-PD cancellation reproducer.