hrm-text : add prefix-LM attention

[bolgacg]

HRM-Text models are trained as prefix LMs: prompt tokens attend to each other in both directions and only generated tokens are causal. This branch already reads prefix_lm into hparams.hrm_prefix_lm; this PR makes the mask honour it, 35 lines across six files.

For these models set_input_kq_mask builds the mask without the causal check, so a token attends to every token of its own sequence in the KV cache or the current ubatch. Generation appends one token per sequence per step, so generated tokens still see only the past, which reproduces the training-time mask. Doing it at the mask level keeps every context type and the server on the same behaviour. Three guards come with it:

• llama_context::decode warns when a batch is split into ubatches (prompt tokens in different ubatches cannot see each other; use -ub at least the longest prompt).
• llama-server disables prompt caching for these models: a cached prefix was computed without the tokens that follow it.
• llama_model_is_prefix_lm() is added next to llama_model_is_diffusion() so tools can check.

Validation against the official transformers implementation (prefix attention), same Q8_0 GGUF, greedy decoding, on DAISY, the 592-question Danish benchmark from the model's own group (server run with --reasoning off so the rendered prompt matches the HF chat template):

run	exact match (%)
this branch, causal only	5.6
this branch plus this PR	8.3
official transformers	8.4
reported in the Mimir paper	9.6

With identical prompt bytes the server's answers are word-identical to transformers on 93% of the 592 questions; the causal build agrees on 29%. On 1,000-token prompts the scores stay equal (65.9 both, 99% of answers identical), and on the group's reading benchmark (MultiWikiQA, 512 rows) the patched build scores 65.8 against the published 66.8. Scripts and logged outputs: github.com/bolgacg/daisy-tools (tools/prefix-run; runs in results/).

Not covered: speculative decoding (draft tokens of one sequence in one batch would see each other).

[noctrex]

Thanks for the submission. We wait for the PR to be accepted and then we can add this also