Harden language model readers against malformed input

[lenzo-ka]

pocketsphinx_lm_convert reads untrusted language model files, and several reader paths crashed on crafted input rather than failing cleanly. Each of the following now rejects the malformed input with a diagnostic and a NULL return, while well-formed models load exactly as before:

• read_dmp_weight_array (ngrams_raw.c) read a weight-array size and per–n-gram weight indices from the file and used them without validation. It now checks the read, rejects a non-positive size, rejects a size that cannot fit in the remaining file before allocating, and bounds-checks each index against the array; failure is propagated through the DMP reader.
• ngrams_raw_free (ngrams_raw.c) dereferenced per-order sub-arrays that are NULL when an earlier order failed to build. It now guards the array and each sub-array before freeing.
• ngrams_raw_read_arpa (ngrams_raw.c) dereferenced the line iterator when an ARPA file ended exactly at the end-mark position. It now checks for NULL and warns instead.
• ngram_model_trie_read_dmp (ngram_model_trie.c) passed a format string with two conversions but one argument when reporting a bad header, reading an argument that was never supplied. It now reports the file name once with a single conversion.
• ngram_model_trie_read_dmp ignored read_word_str's return value, so a truncated word-string block left word_str entries NULL and a later consumer dereferenced them. It now checks the return value and fails cleanly, matching the binary reader.

A new unit test, test_lm_reader_hardening, constructs minimal malformed DMP and ARPA models in the test process and drives each through ngram_model_read, asserting a clean failure without a crash: an invalid weight-array count, a weight-array count larger than the remaining file, a partially built n-gram array freed on an incomplete ARPA section, an ARPA model with no end-mark, and a DMP model with a truncated word-string block. Each case fails against the previous reader implementation.

Reported in #482. The five constructions there reproduce the crashes under AddressSanitizer; the shipped en-us and tidigits models and the DMP fixtures, including a byte-swapped DMP, continue to load with these changes.

Fixes #482

Tested on macOS (Apple clang) and Linux x86-64 (GCC 13.3); full unit and regression suite passes on both.