Imported from GitHub PR openxla/xla#41205

Follows the same pattern as the MemzeroCmd/Memset32Cmd/CublasLtMatmulThunk migrations: eliminate the standalone CuDnnCmd wrapper class by making CuDnnThunk directly implement Command (via TracedCommand) and provide its own Record() method.

Before
------
CuDnnThunk (a plain Thunk subclass) was converted by the command buffer emitter into a heap-allocated CuDnnCmd via Convert<CuDnnThunk>(), splitting the same logical op across two classes with duplicated operand/graph state.

After
-----
CuDnnThunk inherits from TracedCommand (using the protected constructor that preserves Kind::kCuDnn + ThunkInfo), implements Record() directly, and is emplaced into the command sequence via static_cast -- matching GemmThunk, CublasLtMatmulThunk, and MemzeroThunk. CuDnnCmd is deleted entirely.

Record() prefers explicit command-buffer construction when the graph reports SupportsExplicitCommandBufferConstruction()==true (CreateDnnGraphCommand / UpdateDnnGraphCommand) and falls back to RecordTracedCommand otherwise.

Initialize() is made safe for pre-populated graphs (e.g. when tests bypass the fingerprint deserialization path): the short-circuit lives inside absl::call_once to keep the read synchronized with concurrent Initialize() calls from other streams/devices.

Tests
-----
- cuda_command_buffer_thunk_test.cc: existing CommandBufferThunkTest.CuDnnCmd rewired to the new owning-SequentialThunk shape.
- cudnn_thunk_test.cc: new CuDnnThunkCmdBufTest fixture + 4 TEST_F cases covering Record() in {Create, Update} x {explicit, implicit} modes. The implicit path is driven by a test-only FakeDnnGraph that forces SupportsExplicitCommandBufferConstruction()->false and whose Execute() emits a verifiable Memset32 on the trace stream. The Update tests allocate a fresh output buffer to exercise the non-cache-hit path in both UpdateDnnGraphCommand and RecordTracedCommand.

Copybara import of the project:

--
f856b7d88c29d4986b872f74d54b7a56d3ab959f by Shawn Wang <shawnw@nvidia.com>:

[XLA:GPU] Migrate CuDnnCmd into CuDnnThunk

Follows the same pattern as the MemzeroCmd/Memset32Cmd/CublasLtMatmulThunk
migrations: eliminate the standalone CuDnnCmd wrapper class by making
CuDnnThunk directly implement Command (via TracedCommand) and provide its
own Record() method.

Before
------
CuDnnThunk (a plain Thunk subclass) was converted by the command buffer
emitter into a heap-allocated CuDnnCmd via Convert<CuDnnThunk>(), splitting
the same logical op across two classes with duplicated operand/graph state.

After
-----
CuDnnThunk inherits from TracedCommand (using the protected constructor that
preserves Kind::kCuDnn + ThunkInfo), implements Record() directly, and is
emplaced into the command sequence via static_cast -- matching GemmThunk,
CublasLtMatmulThunk, and MemzeroThunk. CuDnnCmd is deleted entirely.

Record() prefers explicit command-buffer construction when the graph
reports SupportsExplicitCommandBufferConstruction()==true
(CreateDnnGraphCommand / UpdateDnnGraphCommand) and falls back to
RecordTracedCommand otherwise.

Initialize() is made safe for pre-populated graphs (e.g. when tests bypass
the fingerprint deserialization path): the short-circuit lives inside
absl::call_once to keep the read synchronized with concurrent Initialize()
calls from other streams/devices.

Tests
-----
- cuda_command_buffer_thunk_test.cc: existing CommandBufferThunkTest.CuDnnCmd
  rewired to the new owning-SequentialThunk shape.
- cudnn_thunk_test.cc: new CuDnnThunkCmdBufTest fixture + 4 TEST_F cases
  covering Record() in {Create, Update} x {explicit, implicit} modes. The
  implicit path is driven by a test-only FakeDnnGraph that forces
  SupportsExplicitCommandBufferConstruction()->false and whose Execute()
  emits a verifiable Memset32 on the trace stream. The Update tests allocate
  a fresh output buffer to exercise the non-cache-hit path in both
  UpdateDnnGraphCommand and RecordTracedCommand.

BUILD
-----
- :command_buffer_cmd drops //xla/stream_executor:dnn (no longer used).
- :cudnn_thunk picks up :command, :traced_command, :command_buffer, :stream,
  //xla:util, absl/log, absl/log:check; drops :errors (no TF_* macros left).
- :cudnn_thunk_test promoted from xla_cc_test to
  xla_test(backends=["gpu"], tags=["cuda-only"]) with the deps needed for
  the new direct-Record tests (cudnn_plugin, cudnn_frontend, command_state,
  collective_params, etc.).
- BUILD integrity audited 1:1 against #include graphs for all modified
  targets.

Merging this change closes #41205

FUTURE_COPYBARA_INTEGRATE_REVIEW=openxla/xla#41205 from shawnwang18:shawnw/migrate_cudnn_cmd_to_thunk f856b7d88c29d4986b872f74d54b7a56d3ab959f
PiperOrigin-RevId: 902579190