Improve run_project feedback and large-cohort scaling

[michaelhallquist]

Summary

• add plain-language status messages before Slurm submission, including a clear notice when BrainGnomes must read a large container file in full
• keep exact container identification in the existing project-local cache, without introducing a user-wide cache
• apply requested subject filters before recursive session discovery to reduce unnecessary filesystem work
• bound subject-selection previews and submission progress output for cohorts with hundreds or thousands of subjects
• update release documentation and tests for the 0.9-2 behavior

This is a follow-up to #76 based on performance review of run_project with the fMRIPrep 25.2.5 pipeline test data.

Validation

• devtools::document()
• full devtools::test(): passed; one opt-in Slurm integration test skipped by its guard
• clean-source devtools::check(): 0 errors, 1 environment warning for unavailable qpdf, and 2 environment/compiler notes
• focused provenance, discovery, progress, dry-run, and subject-ID tests: passed
• synthetic 1,000-subject discovery benchmark: 0.395 seconds for full discovery and 0.073 seconds when selecting one subject

Packaging note

Large local runtime directories and check artifacts were already covered by .Rbuildignore. The final check used a clean source snapshot because R package build enumeration can still traverse a very large live development tree before applying those exclusions.