Add counter-based PTA OOM injection for P/D inference

[yjyang62]

Summary

Restores and improves PTA OOM fault injection so it works during inference on large models (e.g. Qwen-235B) in both Prefill (P) and Decode (D) phases.

The previous branch only incremented g_malloc_call_count in the internal caching allocator malloc() without throwing, so OOM faults could not be triggered. This change introduces a dedicated PtaOomInjector with a shared counter across two injection points:

1. Tensor allocation (NPUCachingAllocator::allocate) — covers prefill and initial memory growth
2. NPU op execution (OpCommand::Run/RunOpApi*) — covers decode and graph-replay inference where allocations are reused

HCCL OOM injection (HCCL_OOM_TRIGGER_COUNT, default 6000) remains unchanged for P2P communication faults between P and D nodes.

Environment variables

Variable	Default	Description
PTA_OOM_TRIGGER_COUNT	30004	Shared PTA event threshold for allocations; set <=0 to disable alloc counting
PTA_OOM_OP_TRIGGER_COUNT	same as above	Optional separate threshold for op events; defaults to PTA_OOM_TRIGGER_COUNT
PTA_OOM_DURING_CAPTURE	0	Set to 1 to allow injection during NPU graph capture
HCCL_OOM_TRIGGER_COUNT	6000	HCCL communication OOM threshold (unchanged)

Usage example (Qwen-235B)

# Prefill node (P) — allocation-heavy phase
export PTA_OOM_TRIGGER_COUNT=30004

# Decode node (D) — op-heavy phase with graph replay
export PTA_OOM_TRIGGER_COUNT=5000   # tune based on warmup op count

# P2P communication fault between P and D
export HCCL_OOM_TRIGGER_COUNT=6000

Files changed

• torch_npu/csrc/core/npu/PtaOomInjector.h — new injector API
• torch_npu/csrc/core/npu/PtaOomInjector.cpp — counter logic and env var parsing
• torch_npu/csrc/core/npu/NPUCachingAllocator.cpp — hook alloc OOM, remove broken counter
• torch_npu/csrc/framework/OpCommand.cpp — hook op OOM on all Run paths