dml: add per-instance mutexes to fix concurrent session crashes

[oysteinkrog]

Description

Adds thread-safety to 4 DML EP data structures that race when multiple InferenceSession instances run concurrently on the same D3D12 device. Without these locks, concurrent DML sessions crash with 0x8000FFFF ("Catastrophic failure") in MLOperatorAuthorImpl.cpp.

Problem

When creating multiple InferenceSession instances that share the same D3D12 device (e.g., running person detection and pose estimation models simultaneously), the DML EP crashes because several internal data structures are not thread-safe:

• BucketizedBufferAllocator::Alloc/FreeResource — concurrent allocations corrupt bucket lists
• CommandQueue methods — concurrent command list submissions race
• ExecutionContext — concurrent Flush/SetCommandRecorder calls race
• DescriptorPool::AllocDescriptors — concurrent descriptor allocation corrupts pool state

Fix

Add per-instance mutexes to each of the 4 classes:

Class	Lock type	Why
BucketizedBufferAllocator	std::mutex + std::atomic for m_defaultRoundingMode	FreeResource releases lock before calling QueueReference to prevent lock-order inversion
CommandQueue	std::recursive_mutex	Re-entrance: ExecuteCommandList → ExecuteCommandLists, Close → GetCurrentCompletionEvent
ExecutionContext	std::recursive_mutex + std::atomic<bool> for m_closed	Re-entrance: Flush ↔ SetCommandRecorder cycle
DescriptorPool	std::mutex	Simple mutual exclusion on AllocDescriptors, Trim, GetTotalCapacity

Each session has its own instances of these objects, so the mutexes only serialize intra-session calls. Cross-session concurrency is fully preserved.

Verification

Tested with concurrent inference stress tests:

• 2 models (person detection + pose estimation) running simultaneously — crashes consistently without fix, stable with fix
• 3 models running simultaneously — stable
• 1000-iteration stress test — no crashes

Tested on NVIDIA GeForce RTX 5070 Ti with DirectML, Windows 11.

Motivation and Context

Applications that run multiple ML models concurrently (e.g., real-time sports analysis with person detection + pose estimation) need concurrent DML sessions for performance. The current code assumes single-threaded access to per-session EP objects, which breaks when sessions share a D3D12 device.

This is a minimal fix — only adding locks where data races were observed. No API changes, no behavioral changes for single-session usage.

[oysteinkrog]

@oysteinkrog please read the following Contributor License Agreement(CLA). If you agree with the CLA, please reply with the following information.

@microsoft-github-policy-service agree [company="{your company}"]

Options:

• (default - no company specified) I have sole ownership of intellectual property rights to my Submissions and I am not making Submissions in the course of work for my employer.

@microsoft-github-policy-service agree

• (when company given) I am making Submissions in the course of work for my employer (or my employer has intellectual property rights in my Submissions by contract or applicable law). I have permission from my employer to make Submissions and enter into this Agreement on behalf of my employer. By signing below, the defined term “You” includes me and my employer.

@microsoft-github-policy-service agree company="Microsoft"

Contributor License Agreement

@microsoft-github-policy-service agree company="Initial Force AS"