[magaonka-amd]

📝 Summary of Changes

Create the DNN (MIOpen) handle before command buffers capture convolutions:

• ConvolutionThunk::Initialize creates the device's DNN handle.
• With more than one device, PlatformUtil::GetStreamExecutors also creates every device's MIOpen handle at
  startup, unless HIP limits capture invalidation to the calling device. A runtime check decides this once per
  process (xla/stream_executor/rocm/rocm_capture_invalidation.{h,cc}).

🎯 Justification

With --xla_gpu_enable_command_buffer=+CONVOLUTION, a convolution's first run is recorded inside a stream capture.
A device without a MIOpen handle creates it there. Creating it calls hipMalloc, which HIP rejects during capture,
and the process aborts. This happens on every GPU except the one that ran autotuning, and on a single GPU when the
executable comes from a compilation cache.

Creating the handle in Initialize keeps it out of the device's own capture. With several GPUs that is not always
enough: HIP without ROCm/rocm-systems#12274 lets the handle's synchronous HIP calls invalidate the captures running
on every other device. On such a HIP, all handles are created at startup, after every device is initialized
(inside each device's Init() it can deadlock in comgr).

The HIP version cannot tell whether openxla#12274 is present: develop reported 7.18 before it merged, and it had been
reverted once before. So the check tests the behaviour. It starts a capture on one device and makes a synchronous
call (hipMemset) on another device from a second thread. It first checks that the same call on the capture's own
device does invalidate the capture. If that control fails, or anything is inconsistent, the handles are created at
startup.

🚀 Kind of Contribution

🐛 Bug Fix

📊 Benchmark (for Performance Improvements)

Not a performance change. The check runs once per process, at client creation. On a HIP without openxla#12274, each
multi-GPU process creates a MIOpen handle per device at startup, also when it has no convolutions.

🧪 Unit Tests:

• convolution_thunk_test: Initialize requests the DNN handle and fails without one.
• rocm_capture_invalidation_test:
  • When the probe fails, every device gets its DNN handle (fake probe, no GPU needed).
  • On a GPU, the probe reports that a same-device call invalidates a capture, and leaves no capture, error or
    current-device change behind.
• platform_util_multigpu_test (2 GPUs): after GetStreamExecutors, device 1 creates its DNN handle while device 0
  records a command buffer, and the recording succeeds.

🧪 Execution Tests:

MI355X (gfx950), ROCm 10.2, through JAX with +CONVOLUTION:

• One convolution that aborted on 2 GPUs, and on 1 GPU from the compilation cache, now passes.
• 34 forward and backward convolution cases pass on 8 GPUs with the stock HIP runtime, with
  fix(hip): implement device-scoped invalidation for stream capture rocm-systems#12274, and with the stock runtime reporting version 7.18.

With an empty MIOpen find database, capture also needs ROCm/rocm-libraries#12265 (or MIOPEN_NAIVE_TIMEOUT=0).