[Bug]: V1 MAX_UTILIZATION paused_requests are never processed in the PP executor loop

System Info

N/A — identified through source inspection of current main (while working on #17882 / PR #18113).

Reproduction

Code inspection of tensorrt_llm/_torch/pyexecutor/py_executor.py:

• With pp_size > 1, the executor always selects _executor_loop_pp.
• _executor_loop_pp only handles the V2 recompute-pause flow (_terminate_recompute_paused_requests / _pause_recompute_paused_requests).
• V1 scheduled_batch.paused_requests — produced by the MAX_UTILIZATION capacity scheduler (both the Python MaxUtilizationPolicy and the C++ bind scheduler, which run with two_step_lookahead under PP) — are neither passed to _terminate_requests nor to _pause_requests anywhere in that loop. The single-rank loops (_executor_loop, _executor_loop_overlap) do both.

Expected behavior

When the V1 MAX_UTILIZATION scheduler decides to pause a started request under pipeline parallelism, the request's resources should be freed and the request reset for replay (req.pause(...)), as in the non-PP loops.

Actual behavior

Under PP, a "paused" request is simply not scheduled that iteration: its KV/resources are never released and its state is never reset, so the preemption the scheduler planned for never materializes. Under sustained memory pressure this can leave the scheduler unable to make room (the same requests get re-paused every iteration while still holding their blocks).

Additional notes

Found while fixing #17882: PR #18113 preserves multimodal payloads for pause replay and threads a for_pause flag through the (PP-)termination path, but the PP loop currently has no V1 pause flow to call it from. Wiring paused_requests handling into _executor_loop_pp touches PP microbatch/inflight-id semantics for all request types, so it is kept out of that PR and tracked here instead.

Start in tensorrt_llm/_torch/pyexecutor/py_executor.py at _executor_loop_pp, then compare its request handling with _executor_loop and _executor_loop_overlap. Trace scheduled_batch.paused_requests for the V1 MAX_UTILIZATION scheduler and the PP microbatch/inflight-id flow. Done means paused requests release resources and reset for replay under PP without breaking existing request types.