[1/N] Bazel build rules for building TRT-MLIR dialect, running dialect lit tests

[sjain-stanford]

[Not ready for review yet]

This includes a hermetic python, and CUDA + TensorRT vendored (inside a docker container).

We do have some more work to get to in a ready-for-review state. In particular, I can think of the following:

• Rebase on the latest main
• Replace some patches (e.g. in mlir-tensorrt/tensorrt/test/lit.cfg.py and mlir-tensorrt/test/lit.cfg.py) with more principled fixes that don't break CMake build
• Try to reuse existing requirements instead of adding a new mlir-tensorrt/requirements.txt
• Reuse existing gpu_tools.py instead of the new mlir-tensorrt/test/gpu_tools.py (this was to mitigate a path issue)
• If possible maybe get mlir-tensorrt/version.bzl to sync with CMake's version here (so both need not be updated on version bumps)

In follow-on PRs (2/N,...) we (Cruise) will be upstreaming convertors from TCP/TOSA -> TensorRT, and more integrations.