[FIX_FOR_VLLM_CUSTOM=ecf9d83520eb217401b47d8a5451a27c5231b8c2] Adapt HPU scheduler, ngram proposer and offloading connector tests to upstream API drift - #1556

…HPU scheduler, ngram proposer and offloading connector tests to upstream API drift (vllm-project#1556)

## Bug 1: Forward throttle_prefills in HPUAsyncScheduler.schedule

- **State machine id**: hpu_async_scheduler_schedule_positional_arg
- **Commit**: 957ba4d

### Root cause
vLLM PR #44558 added a throttle_prefills positional arg to
Scheduler.schedule(); EngineCore calls it positionally but the HPU
override only accepted self.

### Upstream PR
vllm-project/vllm#44558

### Fix
Accept throttle_prefills (default False) on the
HPUAsyncScheduler.schedule override and forward it to
super().schedule().

## Bug 2: Pass num_speculative_tokens to NgramProposer.propose

- **State machine id**: ngram_proposer_propose_missing_positional_arg
- **Commit**: 82155ea

### Root cause
vLLM PR #32374 (Dynamic SD) added a leading num_speculative_tokens
positional arg to NgramProposer.propose().

### Upstream PR
vllm-project/vllm#32374

### Fix
Prepend self.speculative_config.num_speculative_tokens in
propose_ngram_draft_token_ids to match the new upstream signature.

## Bug 3: Align OffloadingConnector stats tests with upstream
flat-metrics API

- **State machine id**: offloading_connector_cpu_to_gpu_metrics_missing
- **Commit**: c1eb9e3

### Root cause
vLLM PR #35669 rewrote OffloadingConnectorStats to a self-describing
{types, data} flat-metric payload, dropping the per-direction
CPU_to_GPU/GPU_to_CPU list shape the tests still asserted.

### Upstream PR
vllm-project/vllm#35669

### Fix
Rewrite test_metrics.py to exercise
increase_counter/observe_histogram/aggregate/reduce/reset against the
new self-describing stats contract.

## Bug 4: Align OffloadingConnector scheduler flush assertions with
upstream defer-on-finish

- **State machine id**: offloading_connector_flush_on_finish_deferred
- **Commit**: 575a178

### Root cause
vLLM commit f428718ffe (PR #45823, "Defer on_request_finished until
in-flight
transfers drain") changed OffloadingConnectorScheduler: a finishing
request with
in-flight store jobs no longer flushes those stores immediately —
finalization
is deferred until transfers drain, and flush now fires only on
preemption or
block reuse. test_concurrent_lookups_of_the_same_prefix and
test_abort_loading_requests still asserted flush-on-finish, so they
failed once
the target vLLM SHA picked up #45823.

### Upstream PR
vllm-project/vllm#45823

### Fix
Drop the stale expected_flushed_gpu_block_indexes assertions in the two
affected
tests (matching upstream's own equivalents, which assert no flush in
these
scenarios). test_request_preemption keeps its flush-on-preemption
assertion,
which upstream still honors.

---------

Signed-off-by: Paweł Olejniczak <pawelx.olejniczak@intel.com>
Signed-off-by: slokesha <slokeshappa@habana.ai>