Wraps the previously top-level test_inference script in a parse_args() +
main(args) + __main__ guard so users can vary the clip, sampling
parameters, model name, seed, device, and dtype without editing the
file. The defaults reproduce the prior hardcoded constants exactly, so
`python src/alpamayo_r1/test_inference.py` is unchanged.

Why this is worth adding:

- Researchers iterating on a single clip want to try several
  --num-traj-samples / --temperature / --top-p settings without
  re-saving the file each time.
- Picking a different --t0-us or --clip-id is the most common
  modification users make and is currently a copy/edit/run loop.
- --seed / --dtype / --device / --model lets the script support quick
  ablations and local checkpoints without changes to source.

New flags (all default to the previous hardcoded values):

    --clip-id                      "030c760c-..." (current example clip)
    --t0-us                        5100000
    --model                        nvidia/Alpamayo-R1-10B
    --num-traj-samples             1
    --top-p                        0.98
    --temperature                  0.6
    --max-generation-length        256
    --seed                         42
    --device                       cuda
    --dtype                        bfloat16   (choices: bfloat16/float16/float32)

Verified locally without GPU/HF by extracting parse_args via AST and
asserting:
- All defaults equal the prior hardcoded values.
- CLI overrides take effect; untouched fields keep defaults.
- argparse rejects invalid --dtype values via choices=.

The trailing "VLA-reasoning models produce nondeterministic outputs..."
note is preserved verbatim so the operational guidance for users with
single-sample runs is unchanged.

Signed-off-by: lonexreb <[email protected]>