fix: normalize float16 embeddings without NaNs

[xblwh]

Normalizing float16 embeddings can produce NaNs for empty or unknown-only inputs because 1e-32 rounds to zero. Computing the norm in float16 also underflows or overflows for small or large nonzero vectors.

Compute float16 norms in float32 while preserving the embedding output dtype. Extend the empty-input regression from #114 across float16/float32/float64 and cover mixed zero/nonzero inputs at different scales with WordPiece, BPE and Unigram tokenizers.

Validation on macOS / Python 3.12:

• The same 36 regression cases give 12 failures on main and all 36 pass with this change.
• coverage run --source=model2vec -m pytest --ignore=tests/integration: 339 passed.
• pytest tests/test_model.py with all CI extras installed: 105 passed.
• pre-commit run --all-files: passed, including Ruff, pydoclint and mypy.
• A real minishlab/potion-base-8M before/after check fixes float16 empty/whitespace outputs; float32, float64 and int8 outputs are unchanged.

[stephantul]

@xblwh has this ever happened to you in a real model? If so, which one?

[xblwh]

I reproduced this locally with minishlab/potion-base-8M, loaded with quantize_to="float16" and normalize=True, while investigating this PR. I do not have a production incident to report.

Here is a minimal example using the exact model revision I tested (macOS / Python 3.12 / NumPy 2.5.3):

import numpy as np
from huggingface_hub import snapshot_download
from model2vec import StaticModel

path = snapshot_download(
    "minishlab/potion-base-8M",
    revision="bf8b056651a2c21b8d2565580b8569da283cab23",
    allow_patterns=["config.json", "tokenizer.json", "model.safetensors"],
)
for dtype in (None, "float16"):
    model = StaticModel.from_pretrained(path, quantize_to=dtype, normalize=True)
    out = model.encode(["", "   ", "Python is a programming language."])
    print(out.dtype, "finite:", np.isfinite(out).all(axis=1).tolist())
    print("empty/whitespace zero:", (out[:2] == 0).all(axis=1).tolist())

On main at d9ac0d9:

float32 finite: [True, True, True]
empty/whitespace zero: [True, True]
float16 finite: [False, False, True]
empty/whitespace zero: [False, False]

With this PR at 69cd5d7:

float32 finite: [True, True, True]
empty/whitespace zero: [True, True]
float16 finite: [True, True, True]
empty/whitespace zero: [True, True]

The real-model failure here is empty/whitespace input after loading in float16; the default float32 model is fine in this example. The very small/large scale cases in the tests are synthetic regression coverage, not failures I observed in this model's normal text embeddings.

[stephantul]

Uh yeah I mean I introduced this regression literally 3 hours ago. It's not even released yet, so unless you installed this today you wouldn't have been able to observe this.

[xblwh]

Right — I checked out today's main at d9ac0d9 while investigating potential issues. That's where I found this and then reproduced it with the Potion checkpoint. I wasn't reporting an issue encountered in a released version. Sorry for not making that timeline clearer.