nsys-jax: re-work to be more pip-install-able

[olupton]

The overarching goal of this PR is to get closer to a world where the nsys-jax tooling is straightforwardly pip install-able. While the diff looks scary, it's mostly re-organisation.

Substantive changes:

• nsys-jax no longer bundles Python code in the output archives, the install.sh script provided for users to run on local machines becomes, loosely, install 'pip nsys-jax[jupyter] @ git+https://github.com/NVIDIA/JAX-Toolbox.git@COMMIT#subdirectory=.github/container/nsys_jax', where COMMIT corresponds to the nsys-jax command that produced the archive. For the ghcr.io/nvidia/jax containers, this is the commit of JAX-Toolbox that triggered the container build.

Changes included:

• Introduce /opt/pip-tools-post-install.d, which pip-finalize.sh will execute the contents of after installing the pip-managed world
  • Migrate install-protoc to use this, so pip-finalize.sh can forget about that detail.
  • Install https://github.com/brendangregg/FlameGraph/blob/master/flamegraph.pl via this.
  • Patch the nvtx_gpu_proj_trace Python code in Nsight Systems 2024.5 and 2024.6 via this.
• Move nsys-jax installation (specifically for the containers) into install-nsys-jax.sh and thereby clean up install-nsight.sh. The new script has to be told the git commit hash of JAX-Toolbox that is being built, because nsys-jax bakes this into an installation script in its output .zip archives to ensure the local environment matches the profile-collection environment.
• The CLI tools like nsys-jax, nsys-jax-combine and install-protoc are now handled via [project.scripts] in pyproject.toml instead of being standalone Python scripts. This is "more standard", and also makes it easier to share code between nsys-jax and nsys-jax-combine.
• The Python library is renamed from jax_nsys to nsys_jax for consistency.
• It's now possible to set the default data loading path via the NSYS_JAX_DEFAULT_PREFIX environment variable; previously the default was the current working directory, but that can be inconvenient to steer in Jupyter environments.