fix(server): sensible default max_ctx + tokenizer auto-detect + Qwen3.6 bench numbers

[dusterbloom]

Three small daemon quality-of-life fixes bundled for review efficiency.

1. Default max_ctx lowered from 131072 → 16384

Related: #10 (the attention-scales-with-max_ctx perf trap).

Current default of 131072 (when DFLASH27B_KV_Q4=1) causes every short-prompt API request to pay a 20×+ slowdown, because the FA kernel strides over the full allocated context instead of the actual filled kv_len. Reproduced after pulling latest main and the daemon-mode commits (#7): a 32K prompt with --kv-q4 --max-ctx=131072 still takes >5 min of prefill vs. 38 s at --max-ctx=32768.

16384 fits typical chat / code / RAG workloads without the trap; users who genuinely need long context can still pass --max-ctx=131072 explicitly.

Help text now points at #10 so anyone surprised can follow the thread.

2. Tokenizer auto-detect

server.py previously hardcoded Qwen/Qwen3.5-27B. Now:

• Reads general.basename / general.name from the target GGUF.
• Maps Qwen3.5-27B → Qwen/Qwen3.5-27B, Qwen3.6-27B → Qwen/Qwen3.6-27B (both validated to share tokenizer.ggml.pre = "qwen35" + identical eos/pad/bos IDs).
• Falls back to the historical default on any error so existing configs don't regress.
• --tokenizer <repo_id> override added for sibling models that land in the qwen35 arch family but aren't in the registry yet.

Startup banner now prints which tokenizer was picked.

3. Qwen3.6-27B bench_llm.py numbers added to README

The Qwen3.6 27b experimental commit (b373478) added prose but limited numbers. This adds the full 3-task bench_llm.py suite on Qwen3.6-27B UD-Q4_K_XL (unsloth Dynamic 2.0, 10 prompts, n_gen=256, RTX 3090 24 GB, auto-fit --max-ctx):

Bench	AR tok/s	DFlash tok/s	AL	Speedup
HumanEval	34.90	78.16	5.94	2.24×
GSM8K	34.89	59.65	4.43	1.71×
Math500	35.13	69.77	5.15	1.99×
Mean	34.97	69.19	5.17	1.98×

For reference, the same harness on Qwen3.5-27B gave 2.87× / 2.21× / 2.56× — cross-generation drop of ~22% uniformly, driven entirely by the draft's hidden-state mismatch against 3.6. Still a clean ~2× with zero retraining.

Test plan

• python3 -c "import ast; ast.parse(open('dflash/scripts/server.py').read())" — parses clean.
• _tokenizer_id_from_gguf() resolves Qwen3.5-27B, Qwen3.6-27B, and bogus paths correctly (unit-tested interactively).
• server.py --help still renders with the new flags.
• End-to-end: DFLASH27B_KV_Q4=1 python3 scripts/server.py --target .../Qwen3.6-27B-UD-Q4_K_XL.gguf boots, serves a request via /v1/chat/completions (tested locally on RTX 3090 before this PR; banner now shows tokenizer = Qwen/Qwen3.6-27B and max_ctx = 16384).