Fix GGUF tensor alignment and validate model config

[dipeshbabu]

This fixes a crash-prone bug in GGUF parsing where tensor data base offset was computed using a bitmask that only works when general.alignment is a non-zero power of two. If alignment is missing/malformed (or non power-of-two), the model can read incorrect mmap offsets and crash.

Changes:

• Use a safe align_up_any() implementation for tensor data base alignment.
• Initialize config with memset to avoid stale/uninitialized fields when metadata is missing.
• Validate required GGUF fields (n_embd/n_heads/n_layers/vocab_size).
• Default n_kv_heads to n_heads when omitted.
• Validate n_embd % n_heads == 0 before computing head_dim.

Tests:

• make clean && make native (Linux x86_64)
• Load and run generation on TinyLlama Q4_K_M gguf

[dipeshbabu]

This also fixes for avoiding per-token malloc/leak by using mmap string views.