Add :safetensors_reader option to load_model/2

[ausimian]

Summary

Adds a :safetensors_reader option to Bumblebee.load_model/2. When supplied, the override is used instead of &Safetensors.read!(&1, lazy: true) to read .safetensors parameter files. The override receives a file path and must return a map from tensor name to an Nx.Tensor or any term implementing Nx.LazyContainer — the same shape Safetensors.read!/2 already returns, so the rest of the loading pipeline (Bumblebee.Conversion.PyTorchParams.load_params!/4) is unchanged.

Default behaviour is identical when the option is not supplied.

Motivation

This is a small seam that enables custom safetensors readers without forking Bumblebee. The concrete use case is a memory-mapped reader backed by a NIF resource binary (enif_make_resource_binary), which keeps peak BEAM memory bounded to a single tensor when loading very large checkpoints — pages are demand-faulted by the OS and freed after the backend transfer, instead of being pread-ed and copied into BEAM heap per tensor.

The current Safetensors.read!/2 lazy path already streams reasonably well, but the per-tensor File.open + pread + binary copy is a real cost on multi-GB checkpoints, and there's no way to swap it today because the loader is wired directly to the module name at bumblebee.ex:771.

The option is scoped to safetensors only. PyTorch pickle isn't a candidate for the same treatment because it requires structural parsing rather than byte-range reads.

Test plan

• mix test test/bumblebee_test.exs --only describe:"load_model/2" passes locally (5/5).
• New test asserts the override is invoked for .safetensors files and that the resulting params have the same keys as the default reader.
• Existing tests pass unchanged (default path is unaffected).