Adding Telemetry support for GPU EP

[urpetkov-amd]

GPU EP Telemetry

Lightweight, fire-and-forget telemetry for the AMD GPU Execution Provider. Each
model load appends a single line to a shared log file describing the EP version,
backend, GPU architecture, model, and cache state.

Goals

• Lightweight: one line per model load, no measurable runtime cost.
• Safe under concurrency: many processes/threads may write the same file.
  Each append takes an exclusive lock, writes, and releases.
• Never affects the caller: if the file can't be opened, or the lock can't
  be acquired within ~10 ms, the record is silently dropped. Telemetry never
  throws and never propagates a failure into inference.

Log location

Platform	Path
Windows	%ProgramData%\AMD\GPUEP\telemetry.log
Linux	/var/log/AMD/GPUEP/telemetry.log (override with AMD_GPUEP_TELEMETRY_DIR)

%ProgramData% is machine-wide and writable by normal (non-elevated) user
processes, which suits inference hosts. A compile-time switch
(telemetry::kUseProgramFiles in telemetry.h) can move the log to
%ProgramFiles%\AMD\GPUEP\ once the installer grants write access there via an
ACL — until then, %ProgramFiles% would silently fail for non-elevated
processes (by design).

Record format

One newline-terminated line per model load:

<ISO-8601-UTC> v=<schema> ep_ver=<v> backend=<name> gfx=<arch> model=<file> mxr_cache=<hit|miss> parent=<proc>

Example:

2026-06-19T13:50:03Z v=1 ep_ver=0.1.0 backend=MIGraphX gfx=gfx1100 model=synthetic_model.onnx mxr_cache=miss parent=python.exe

• key=value was chosen over CSV/JSON so the consumer can add or ignore fields
  without breaking parsers.
• A leading v=<schema> (currently 1) lets the format evolve.
• Optional fields are omitted when unknown (e.g. gfx/mxr_cache are not
  emitted on the DirectML path).
• Values are sanitized (whitespace/newlines → _) so a value can never split a
  line.

Telemetry component (src/shared/common/)

Follows the repo's existing platform-split convention:

File	Responsibility
telemetry.h	Public API: Record, Log(), AppendLine(), LogFilePath(), process-name helpers, constants
telemetry.cc	Cross-platform: Record::Format(), LogFilePath() (path join + create_directories), Log() orchestration
platform/windows/telemetry.cc	LockFileEx append, %ProgramData% from env, Toolhelp process lookup
platform/linux/telemetry.cc	fcntl(F_SETLK) append, base dir, /proc process lookup

All cross-cutting logic lives in telemetry.cc; the platform files contain only
the irreducible OS-specific code.

Locking mechanism

Exclusive file lock with a ~10 ms acquire timeout. On
Windows this maps directly to LockFileEx:

1. Open with GENERIC_READ | GENERIC_WRITE, FILE_SHARE_READ | FILE_SHARE_WRITE,
   OPEN_ALWAYS. Sharing is allowed; the byte-range lock provides mutual
   exclusion.
2. Acquire the lock with LOCKFILE_EXCLUSIVE_LOCK | LOCKFILE_FAIL_IMMEDIATELY in
   a poll loop bounded by GetTickCount64() + ~10 ms (LockFileEx has no native
   timeout). Give up silently on timeout.
3. Seek to EOF under the lock and WriteFile.
4. UnlockFileEx, CloseHandle.

A sentinel byte far past EOF (offset 0x7FFFFFFF'FFFFFFFF) is locked as the
gate rather than byte 0, so Windows' mandatory locks never block a concurrent
reader.

Files changed

New:

• src/shared/common/telemetry.h
• src/shared/common/telemetry.cc
• src/shared/common/platform/windows/telemetry.cc
• src/shared/common/platform/linux/telemetry.cc
• src/shared/tests/telemetry_stress_test.cc

Modified:

• src/shared/CMakeLists.txt — add telemetry sources; add BUILD_TELEMETRY_STRESS_TEST option.
• src/amdgpu/gpu_ep.{h,cc} — record backend choice; emit telemetry for the
  DirectML path (LogTelemetry, std::call_once).
• src/amdgpu/gpu_factory.h — public Version() accessor.
• src/migraphx/mgx_ep.{h,cc} — emit full telemetry record in Compile();
  thread MXR cache-hit out of CreateNodeComputeInfoFromGraph.
• src/migraphx/mgx_factory.h — public Version() accessor.
• src/amdgpu/pyproject.toml — package the migraphx-backend and
  directml-backend components (see Related change).