Bump llama.cpp to a646006 (b9932) for v0.8.34

[nyo16]

Summary

• Bumps vendor/llama.cpp from cb295bf59 to a646006f0 (44 commits, tag
  b9932) and bumps the package version to v0.8.34.
• No NIF changes required — every header the binding compiles against
  changed additively only: new Q2_0 quant enum values in ggml.h/llama.h, a
  new common_speculative_init_result helper in common/speculative.h, and a
  server_base field + <fstream> include in common/common.h.
• Adds speculative-decoding support-status docs (README + LlamaCppEx.MTP
  moduledoc): the binding exposes MTP only; DFlash/EAGLE-3/n-gram exist upstream
  behind the same common_speculative API but are not reachable through
  speculative_init yet.

DFlash on Apple Silicon (tested at b9932, M4 Max)

Ran upstream llama-cli built from this exact submodule commit with
--spec-type draft-dflash, Qwen3.5-4B Q4_K_XL target + z-lab 0.6B drafter
(f16):

Config	Gen t/s	Acceptance
plain (greedy)	85.1	—
DFlash (greedy, n-max 6)	42.0	30% (mean run 2.8)
DFlash (temp 0.8, n-max 15)	27.3	6.7%

Functional end-to-end on Metal, but net slower at this target size — same
wide-verify-batch economics as the MTP note in the README. Community
drafter-GGUF conversions are also still incompatible across converter
revisions (upstream
#25116 arch mismatch,
#25110 metadata refactor);
of three third-party drafter repos tested, only one loads with current
upstream. Documented in the README rather than exposed as a feature.

Testing

• Rebuilt the test-env NIF from scratch against b9932 (Metal): mix test --include smoke with local gen/embedding models → 191 passed, 0 failures, 4
  skipped.

checksum.exs intentionally untouched — CI regenerates it after the release
tag.