gguf : align the data section relative to the GGUF start, not the file

[ykhrustalev]

Overview

Rework of #28973 using FILE * instead of file descriptors, as asked there and in #20402.
#28973 can be closed.

gguf_init_from_file_ptr() reads from the current file position, but aligns the data
section from offset 0. So a GGUF stored at an unaligned offset inside a bigger file (an
Android APK asset, for example) loads without error and returns wrong tensor data. The
reader now aligns from where the GGUF starts.

With that fixed, no new API is needed for embedded models: fseek to the GGUF, then
llama_model_load_from_file_ptr(). On Android:

FILE * f = fdopen(dup(fd), "rb");      // fd, start, length from AssetFileDescriptor
fseeko(f, start, SEEK_SET);
llama_model * model = llama_model_load_from_file_ptr(f, params);
fclose(f);

Two small additions:

• llama_adapter_lora_init_from_file_ptr(), since LoRA had no FILE * entry point.
• If the GGUF data is not on a 32 byte file offset, mmap is turned off with a warning
  instead of ggml asserting.

Additional information

• test-gguf: new file_offset mode, GGUF written after 7 junk bytes. Fails without
  the fix.
• test-load-file-ptr (needs the model fixture): model and LoRA loaded from an offset,
  logits compared with a normal load. It is a new test file; drop it if not wanted.
• Passes on macOS arm64 and a Galaxy S24 Ultra, and from a real APK on the S24 where the
  LoRA asset landed at offset mod 32 = 16.

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: YES. An assistant helped find the bug, write the fix and tests, and
  run the checks.

[JohannesGaessler]

This should throw a runtime error instead.