Add run_nmf_batch(), opt-in parallel NMF, and g:Profiler hardening

[MilesGithub]

Batch-and-scale features. Everything is additive; the sequential default is unchanged.

Changes

• run_nmf_batch() — runs NMFprofileR() over a named list of cohorts, each to its own <name>_Results/ subdirectory, and returns one consolidated per-factor summary tagged by cohort Run_ID. It is resilient (on_error = "continue" records failures and carries on), resumable (skip_existing reloads finished cohorts from their .rds bundle or summary), and writes Batch_Consolidated_Summary.tsv.
• Opt-in parallel NMF — nmf_parallel = TRUE (with optional nmf_cores) spreads the consensus runs across a local cluster. NMF selects its worker packages from the attached search path, so nmf_fit() attaches NMF for the duration of a parallel fit (via attachNamespace(), restored afterwards); otherwise workers hit the old "none of the packages are loaded" failure when NMF is called from a package namespace. Verified that for a fixed seed the parallel result is byte-identical to the sequential path (basis, coef, objective). Default remains sequential.
• g:Profiler hardening — an oversized-query guard (max_query_size), automatic retry with exponential backoff on transient failures, and an optional on-disk result cache (enrichment_cache) keyed by a stable query hash so re-runs and batches skip the network. Exposed on NMFprofileR() and nmf_enrichment().

Testing

• New network-free tests: test-batch (consolidation, Run_ID tagging, skip_existing, on_error, arg guards, via a mocked fit), test-enrichment-hardening (hash key, retry, oversized guard, cache hit), and test-parallel (parallel reproduces sequential; asserts it did not fall back).
• Repointed the stale determinism parallel test at the working nmf_fit() path (it previously always skipped).
• Full suite green on R 4.6.0 (0 failures/0 warnings). R CMD check clean (0 errors/warnings/notes), both vignettes build. New "Batch analysis across many cohorts" vignette. Version 0.4.0.