chore(ci/litjob): fix and optimize GPU related tests

[bhimrazy]

What does this PR do?

This PR has the following changes:

• Reduced GPU test time by approximately 25–30 minutes.
  Two of these tests were taking more than 20 minutes but, when run manually, finished within 5 seconds.

  • test_tensorrt_save_ir_type[ts-ScriptModule]
  • test_tensorrt_export_reload[ts-torchscript]
    

  The delay in those tests was mainly caused by CUDA_LAUNCH_BLOCKING=1.

  Before:
  
  After:
  

• Fixes a typo by replacing $(PACKAGE_NAME) with ${PACKAGE_NAME}.

• Fix: unzip not found — used in ci

  .actions/pull_legacy_checkpoints.sh: line 10: unzip: command not found
  total 0
  + cd tests/legacy
  + ls -lh checkpoints/
  total 0
  

  pytorch-lightning/.actions/pull_legacy_checkpoints.sh

  Line 10 in 573828a

  unzip -o tests/legacy/checkpoints.zip -d tests/legacy/

Note: It seems there are still ci/litjob failures caused by port issues, which require a rerun to pass.

PR review

Anyone in the community is welcome to review the PR.
Before you start reviewing, make sure you have read the review guidelines. In short, see the following bullet-list:

Reviewer checklist

• Is this pull request ready for review? (if not, please submit in draft mode)
• Check that all items from Before submitting are resolved
• Make sure the title is self-explanatory and the description concisely explains the PR
• Add labels and milestones (and optionally projects) to the PR so it can be classified

[bhimrazy]

It seems the delay in those tests is caused by CUDA_LAUNCH_BLOCKING=1.

Test Code

export CUDA_LAUNCH_BLOCKING=0
time python -c "
import torch
from lightning.pytorch.demos.boring_classes import BoringModel
model = BoringModel()
model.example_input_array = torch.randn((4, 32))
model.cuda()
print('Starting compilation with CI env vars...')
import time
start = time.time()
ret = model.to_tensorrt(ir='ts')
end = time.time()
print(f'Compilation took {end - start:.2f} seconds')
"

Almost 20 mins saved