Add support for Talkie (talkie-1930-13b)

[ethanelasky]

Adds the Talkie architecture — talkie-lm's 13B models trained entirely on pre-1931 text (model card). llama.cpp already supports it as LLM_ARCH_TALKIE. The official repos are raw PyTorch checkpoints, but the transformers port at lewtun/talkie-1930-13b-it-hf converts directly with mlx_lm.convert and ships a normal tokenizer.json + chat template, so this is just a model file and a test entry. No tokenizer changes, no new dependencies.

Mostly a standard 40-layer MHA decoder (40 heads x 128, SwiGLU, vocab 65540), with a few quirks:

• every RMSNorm is weightless (no learned scale), fp32 eps
• RoPE rotates with the opposite sign to NeoX. Implemented with mx.fast.rope and negated freqs, same trick as nanochat.py; checked against the PyTorch reference (max diff ~5e-7). llama.cpp gets the same effect by folding the sign flip into the q/k weights at conversion time.
• per-head RMSNorm on q and k applied after RoPE, plus a learned per-head gain on q only
• learned scalar gains on the attention and MLP residuals, and a per-block skip connection from the normalized token embeddings
• lm_head is a bare tensor scaled by a learned scalar (what the GGUFs export as logit_scale)

The offset goes straight from the cache into mx.fast.rope (int or per-sequence array) and there's no custom make_cache, so batch_generate and the server's continuous batching work as-is.

Re: the two earlier talkie PRs (#1220, #1231) — both build RoPE from a precomputed cos/sin table indexed by a scalar cache[0].offset and pin make_cache to KVCache, so they don't work with BatchKVCache. #1220 also targets the original checkpoint layout rather than the -hf one, and adds a tiktoken-based tokenizer adapter plus a tiktoken dependency that turns out to be unnecessary: the -hf repo's tokenizer.json tokenizes identically to llama.cpp's GGUF tokenizer on everything I threw at it, special tokens included.

Testing:

• converted lewtun/talkie-1930-13b-it-hf to bf16 and q8 (group size 64); both load strictly with no missing/unexpected keys
• q8 vs bf16 last-position logits on 12 prompts: argmax matches 12/12, max logit delta under 0.9 on 11 of them (one prompt reaches 3.7 with argmax unchanged — this model has unusually large activation outliers and some prompts are sensitive)
• greedy decode vs llama.cpp (CPU, Q8_0): exact 64-token agreement on 4/12 prompts; every divergence is a near-tie between the two different q8 grids (our token sits at rank 2 in llama.cpp's logprobs)
• batched generation matches single-stream greedy at prompt lengths 1/8/9/500 in one batch, apart from single-token flips at exact bf16 ties. Measured with Fix float32 promotion in BatchKVCache/BatchRotatingKVCache extend() #1491 applied — without that fix, sequences joining an in-flight batch get fp32-promoted K/V, which this model is sensitive to. Not a dependency, just affects that particular comparison.
• added a talkie entry to test_all_models; full tests/test_models.py passes and pre-commit is clean

To try it:

mlx_lm.convert --hf-path lewtun/talkie-1930-13b-it-hf --mlx-path talkie-q8 \
    -q --q-bits 8 --q-group-size 64
mlx_lm.generate --model talkie-q8 --temp 0 \
    --prompt "What is the proper way to address a letter to a duchess?"

which answers: "To the Duchess of —, Belgrave Square, London, W."