No lazy/mmap loading for GGUF weights (MmapedSafetensors equivalent)

Finding from embedding candle's quantized GGUF models in a long-running server process, verified against the 0.11 sources.

`gguf_file::TensorInfo::read` does `reader.read_exact(&mut raw_data)` per tensor — loading a GGUF always copies the full file into RAM (`from_gguf` → `Content::tensor` → `TensorInfo::read`, candle-core `quantized/gguf_file.rs`). safetensors has `MmapedSafetensors` for lazy loading, but there is no GGUF equivalent, so wrapping the reader over `memmap2` only changes the copy mechanism — it cannot avoid the copy.

An mmap-backed `QTensor` storage for GGUF would:

- cut cold-start load time (measured: ~6s for a 0.5B Q4_K_M, ~16-19s for a 3B on NVMe — essentially all of it read+copy);
- let the OS page cache share weights across processes running the same model;
- reduce peak RSS during load (today the copy briefly doubles resident memory pressure on small machines).

Start in candle-core/quantized/gguf_file.rs and trace the mentioned from_gguf → Content::tensor → TensorInfo::read path. Compare the existing safetensors MmapedSafetensors approach, then investigate how an mmap-backed QTensor storage could avoid copying each tensor. Done means GGUF weights load lazily without copying the full file into RAM.

Written by the indexing model from the issue text.