Add LoRA analysis: SVD spectrum, effective rank, Frobenius norms

[harelc]

• Add LoRA pair detection (lora.py) with support for common naming conventions (lora_A/B, lora_down/up, down/up)
• Add tensor data loading and SVD computation (data.py) with QR-based optimization to avoid full-matrix SVD
• Add LoraScreen with DataTable listing all LoRA pairs, columns for rank, ||A||, ||B|| Frobenius norms, and effective (stable) rank
• Add SvdScreen with full-width vertical bar histogram (sqrt scale) showing singular value spectrum with effective rank statistics
• Add LoraHelpScreen (?) explaining stable rank, Frobenius norm, and SVD spectrum metrics
• Context-aware L key: filters pairs to current view, jumps directly to SVD when only one pair matches
• Async stats computation with 4-thread pool: norms first (cheap), then stable rank (SVD-based), cells update progressively
• Export (on user request) and load (automatically) the LoRA statistics analysis as it's computationally heavy.

[matanby]

Hi @harelc — thanks a lot for putting this together, the QR-SVD trick and the per-pair stats screen were both great ideas. I ended up reimplementing the core pieces against the new multi-command structure rather than merging the branch directly, but most of the work is now in v0.2.0:

What landed (inspired by this PR):

• QR-accelerated SVD — src/sft/ops/lora/svd.py, the same qa, ra = qr(a.T); qb, rb = qr(b); svd(rb @ ra) trick. Used by sft lora svd, sft lora resize, and sft lora stack --target-rank.
• LoRA Mode in the TUI (L key) — LoraModeScreen with per-pair rank / ||A|| / ||B|| / effective rank columns; Enter on a row opens the singular-value spectrum modal (SvdSpectrumScreen).
• Async stats worker — norms fill in first, stable rank is computed in the background, cells update progressively.
• PEFT + Kohya pair detection, plus a sft lora convert command to move between them.

Not (yet) backported, in case you want to follow up:

• Bare down/up naming (SDXL-style without lora_ prefix) — detect.py currently recognizes PEFT (lora_A/lora_B) and Kohya (lora_down/lora_up); the third convention from your PR isn't covered.
• Dedicated ? help modal explaining stable rank / Frobenius norm / SVD — we currently surface those definitions in command --help text and the Compress prompt, not as a modal inside LoRA Mode.
• Persisted analysis cache (your export/auto-load idea) — with the QR-SVD path the recompute is fast enough that we skipped this, but it would still be nice for very large adapters.
• Context-aware L filtering to the current view — L always opens the full LoRA list right now.

Closing this PR since the core features are in v0.2.0; happy to take a follow-up PR for any of the gaps above. Thanks again!