Add a CPU-first TraceML demo container

[sahilmathur254]

What changed

• Add a root Dockerfile based on Python 3.10 slim.
• Install TraceML from the checked-out source with .[torch].
• Run the slow DataLoader bottleneck demo in summary mode by default.
• Exclude repository metadata and local or generated artifacts from the image.
• Document the CPU-first build and run commands, with GPU access presented only as optional guidance for configured NVIDIA environments.

Why

TraceML did not have a container-based first-run path. This gives new users a single-image demo that requires no host Python environment, external dataset, or GPU and prints the input-bound diagnosis requested in #216.

How I tested

• Tests added or updated
• Existing tests run
• Manual smoke test
• Not run; reason:

Validation performed in an isolated Docker context:

• docker build -t traceml-demo .
• docker run --rm traceml-demo
  • exited with status 0
  • ran on Device: cpu
  • completed 60 training steps
  • printed TraceML Verdict: INPUT-BOUND / CRITICAL
• Verified the image excludes Git metadata, OS files, Python caches, virtual environments, generated logs, and pip build artifacts.
• git diff --check

No unit tests were added or run because the change is limited to the container entry point, build context, and documentation; the issue's end-to-end Docker acceptance path was exercised directly.

Runtime impact

• No training-path impact
• May affect tracing, runtime, telemetry, or distributed behavior
• Not sure

Notes: The change only adds an opt-in container entry point and documentation. The locally resolved image is about 3 GB because the current Linux PyTorch dependency set is substantial.

Docs

• Updated
• Not needed

Extra context

The default command follows the maintainer-confirmed CPU-first scope:

traceml run examples/diagnosis/dataloader_bottleneck_demo.py --mode=summary --args --scenario slow

The optional --gpus all example does not guarantee CUDA support; it is conditional on NVIDIA Container Toolkit and a CUDA-capable PyTorch installation. GPU execution was not tested and is not an acceptance requirement.

Closes #216