Add in-memory model loading API

[jolestar]

This PR adds a small, self-contained loading improvement that makes Model2Vec easier to embed in runtimes that already fetch model assets as bytes.

What changes

• adds StaticModel::from_bytes(...) for loading a model from in-memory tokenizer.json, model.safetensors, and config.json bytes
• extracts resolve_model_files(...) so path / Hugging Face file resolution is reusable
• rewrites from_pretrained(...) to reuse the same byte-loading path instead of duplicating model parsing logic
• adds a regression test that verifies from_bytes(...) matches local from_pretrained(...)

Why

This keeps the existing API intact, but makes it possible to integrate model2vec-rs into environments where model assets are fetched or embedded outside a local filesystem-oriented flow.

I kept this PR intentionally narrow and did not include any feature-flag or wasm-specific dependency changes yet.

Validation

• cargo test

[Copilot]

Pull request overview

This PR adds an in-memory loading entrypoint for StaticModel and refactors the existing from_pretrained path to reuse a shared byte-parsing implementation, improving embeddability in environments that already provide model assets as bytes.

Changes:

• Added StaticModel::from_bytes(...) to construct a model from in-memory tokenizer.json, model.safetensors, and config.json bytes.
• Extracted resolve_model_files(...) to centralize local-path vs Hugging Face Hub file resolution and updated from_pretrained(...) to reuse it.
• Added a regression test ensuring from_bytes(...) matches from_pretrained(...) for a local fixture model.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 3 comments.

File	Description
tests/test_model.rs	Adds a regression test comparing embeddings produced by byte-loaded vs path-loaded models.
src/model.rs	Introduces from_bytes, refactors from_pretrained to read bytes and delegate, and adds reusable model-file resolution helper/types.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

A module-level use std::fs; was added, but test_from_borrowed already imports std::fs inside the function. This duplication is easy to miss and can trigger lint noise; consider removing the inner import and using the module-level one consistently (or vice versa).

[Copilot]

ModelFiles is a new public struct but it has no rustdoc and exposes its fields as pub, which locks in the exact file naming/layout as part of the public API. Consider making this pub(crate) (if only intended for internal reuse) or adding rustdoc + keeping fields private (with accessors) to preserve API flexibility.

Suggested change

	pub struct ModelFiles {
	pub tokenizer: std::path::PathBuf,
	pub model: std::path::PathBuf,
	pub config: std::path::PathBuf,
	}
	/// Resolved file paths for a Model2Vec model.
	///
	/// This struct groups together the paths to the tokenizer, model weights, and
	/// configuration files, regardless of whether they were loaded from a local
	/// directory or downloaded from the Hugging Face Hub.
	pub struct ModelFiles {
	tokenizer: std::path::PathBuf,
	model: std::path::PathBuf,
	config: std::path::PathBuf,
	}
	impl ModelFiles {
	/// Returns the path to the tokenizer JSON file.
	pub fn tokenizer(&self) -> &std::path::Path {
	&self.tokenizer
	}
	/// Returns the path to the model weights file (in safetensors format).
	pub fn model(&self) -> &std::path::Path {
	&self.model
	}
	/// Returns the path to the model configuration JSON file.
	pub fn config(&self) -> &std::path::Path {
	&self.config
	}
	}

[Copilot]

resolve_model_files is now pub and has a global side effect (env::set_var("HF_HUB_TOKEN", ...)). As a reusable API this can create surprising cross-call behavior (e.g., concurrent callers, different tokens, test isolation). Prefer constructing the hf-hub client with the token (e.g., an ApiBuilder/token option) or keep the env-var mutation confined to from_pretrained and make this helper pub(crate).

[jolestar]

Followed up on the initial review pass: the helper pieces introduced in this PR are now kept internal, and I also cleaned up the duplicated test import. The PR remains intentionally narrow: it adds in-memory model loading plus the loading-path refactor, and I re-ran cargo test after the cleanup.

[Pringled]

Hey @jolestar, thanks for adding this, LGTM. CI is failing, can you run clippy to fix the formatting? After that it's good to go.