Add conflict-resolution modes (norm-scaler, gram-schmidt) to lora add and lora stack

[IsraelBenDavid]

What

Adds a --mode option to both sft lora add and sft lora stack for resolving
conflict between adapters before they're combined. Two strategies:

• norm-scaler — rescale the other adapter's delta so its Frobenius norm
  matches the reference's (‖ΔW_other‖_F == ‖ΔW_ref‖_F), so neither update
  dominates purely by magnitude.
• gram-schmidt — subtract the reference-aligned component of the other
  delta (ΔW_other − c·ΔW_ref, c = ⟨ΔW_ref,ΔW_other⟩/⟨ΔW_ref,ΔW_ref⟩), so the
  added adapter only contributes directions the reference doesn't already cover.

--mode none (default) preserves the existing behavior.

Why

Naively summing LoRA deltas lets one adapter overpower another by magnitude, or
double-counts overlapping directions. These transforms give a principled way to
combine adapters that "fight" each other.

How

• New module conflict.py holds the math:
  norm_scale_factor, gram_schmidt_orthogonalize, the factored Frobenius inner
  product frob_inner_factored, and validate_mode.
• add (lossy): reconstructs each module's delta, transforms every
  non-reference delta against adapter 0, then does the usual weighted sum + SVD
  re-decomposition.
• stack (lossless): because both transforms are linear, they fold into the
  concatenated factors as a per-module coefficient tweak — the merge stays exact
  at rank r_a + r_b. Scalars are computed in rank space, so the full out×in
  delta is never materialized.
• The reference adapter is the first file for add and file A for
  stack. The chosen mode is recorded in output metadata (conflict_mode).

Tests

New unit + CLI coverage in tests/test_lora_add.py and
tests/test_lora_stack.py: the math primitives,
zero-delta / single-side no-ops, that stack modes stay lossless and produce the
expected norm/orthogonality, metadata recording, invalid-mode rejection, and the
CLI wiring for both commands.

Docs

Updated SKILL.md and REFERENCE.md with the new option and examples, plus a
note clarifying add is lossy (fixed output rank) vs stack being lossless.