migrate internal changes

[christopherbate]

• [mlir-tensorrt] NFC: fix build when NCCL is ON but MPI is OFF
  GitOrigin-RevId: 8c365f7839bc39903c21a6fffe370bcff0f3f074

• [mlir-tensorrt][cmake] Refactor compilation options into dedicated CMake module, add extra checks mode
  This change consolidates handling of global compilation options (that are
  specific to MTRT and not handled as part of HandleLLVMOptions) into
  a dedicated CMake module. It adds additional CMake cache options for
  enabling '-Werror' globally and for addition additional warnings and
  runtime checks recommended by
  https://best.openssf.org/Compiler-Hardening-Guides/Compiler-Options-Hardening-Guide-for-C-and-C++.html.

  By default, MLIR_TRT_ENABLE_EXTRA_CHECKS and MLIR_TRT_ENABLE_WERROR are
  set to OFF, but the idea is to enable them in CI in the future.

  In addition, fix some lingering warnings and build issues discovered
  when testing with the extra checks mode and Werror enabled:

  • Supress deprecated declaration warnings related to TensorRT
  • Fix some warnings when NCCL is enabled.
  • Fix test flags when ORTools is OFF but Shardy is ON.

  GitOrigin-RevId: 39f5e91697a20a65da2a5374b90366aba005f889

• [mlir-tensorrt] Fix const fold execution aliasing and multi-device runtime issues
  Fixes two issues that manifested as non-fatal error logs related to
  CUDA runtime function call failures during compilation or execution
  of a test program. Such errors were logged when e.g. we have no
  path to propgate the error outwards (such as calling cudaFree in a
  destructor).

  Despite the similar symptoms, there were too different root causes:

  1. In the execute-constant-foldable-subgraphs pass, we previously
     had an issue where a function returning multiple results could return
     two values which alias each othera, which is a violation of the
     compiler-runtime ABI contract. We fixed this for ABI v1 in
     ee26682 but forgot to update the
     actual pipeline in the "execute-constant-foldable-subgraphs" pass.
     This is now fixed. In addition, we now give each outlined sub-module
     a unique name so that the IR printing doesn't clobber debug artifacts
     for different constant foldable modules.

  2. The second issues is that in multi-device setups, we need to ensure
     that CUDA events created as part of the pinned memory pool allocator
     are associated with the correct device. Simplify the existing infra
     by moving the pinned memory allocator to be part of the runtime client
     only, not the runtime session. The allocator now has a CUDA event
     pool for each device.

  GitOrigin-RevId: 2c5b5e7026d04150ebae2f40e25faf42d2af70df

• [mlir-tensorrt] NFC: bump version to 0.4.3
  GitOrigin-RevId: 9eb762aaf5edcc4e24220a67cc96c25b94c36173

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.