docs : add TurboQuant project overview to AGENTS.md and skills

[giveen]

Overview

• AGENTS.md: fork overview (turbo/TQ types, key files, env knobs, test gates, git workflow, known pitfalls)
• skills/code-review: fork-specific TurboQuant checklist
• skills/add-new-model: fork-specific validation and dedup steps
• remove REBASE_PLAN.md (planning doc, superseded)

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: Yes

[TheTom]

Merging. I fact-checked the claims rather than skimming, because a wrong line in an orientation doc propagates further than a wrong line in a function: it gets believed by everyone who arrives afterwards, including agents.

Everything checkable, checked against the tree:

claim	verdict
TURBO2_0 = 43, TURBO3_0 = 44, TQ3_1S = 45, TQ4_1S = 46, TURBO4_0 = 47	correct, including TURBO4_0 at 47 rather than the intuitive 45
turbo types are runtime-only, never in GGUF	correct
TQ3_1S/TQ4_1S have CPU, CUDA/HIP, Metal, Vulkan and SYCL kernels	correct, SYCL included (25 matching sources)
docs/KV-cache-quantization.md is the authoritative usage doc	exists
ggml/src/ggml-cuda/mmvq-tq.cu native TQ kernels	exists
MLA models skip V rotation, K and V types must match	consistent with the is_mla handling in llama-kv-cache.cpp

The type table getting TURBO4_0 = 47 right is the detail I would expect a hand-written doc to get wrong, since the sequence looks like it should be contiguous. That it survives upstream's Q2_0 claiming 42 is exactly the kind of thing a newcomer needs told.

Removing REBASE_PLAN.md is right; it described a migration that has since happened, and stale plans are worse than absent ones.

One suggestion for a follow-up rather than a change request: the doc would be more useful with a short section on what the test suites do and do not cover, because that has been this fork's most expensive recurring surprise. Three separate times in the past week a green run turned out to be measuring nothing: MUL_MAT_ID passing at n=16 while the decode path was never exercised, SET_ROWS TQ4_1S cases passing vacuously on zero rows, and test-backend-ops reporting OK for an op whose every case was skipped as unsupported (#242). A paragraph saying "a passing suite means the cases that ran passed, check that yours ran" would save the next person the same three discoveries.

Thanks for writing this down. Orientation docs are unglamorous and this one is accurate.