Add float16 and bfloat16 support in sinusoidal encodings.

[gdewael]

Currently, the transformers in depthcharge should not be trained with float16 or bfloat16 precision, which has become semi-standard in transformer training.
The issue is two-fold: self.sin_term and self.cos_term have large numbers which cannot be accurately represented with lower precision, distorting the positional encodings (PositionalEncoder(1024).sin_term.max() returns tensor(15915.4941)), and the mz resolution of MS experiments can not be faithfully represented using bfloat16 and float16 inputs (i.e., mz_array inputs to SpectrumTransformerEncoder, see #76 (comment)).

This PR does the following:

• When FloatEncoder's learnable_wavelengths=False, it fixes self.sin_term and self.cos_term to float32 precision.
• When FloatEncoder is passed lower than float32 precision values, it raises a warning.
• It introduces a dummy buffer to convert the sinusoidal encodings to the proper model precision after they have been computed.
• A similar procedure is followed for PositionalEncoder, where this line would otherwise give inaccurate positions for large sequences in combination with bfloat16 model dtypes.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 96.69%. Comparing base (5a1afb1) to head (caf27e8).

Additional details and impacted files

@@            Coverage Diff             @@
##             main      #82      +/-   ##
==========================================
+ Coverage   96.62%   96.69%   +0.06%     
==========================================
  Files          24       24              
  Lines        1008     1029      +21     
==========================================
+ Hits          974      995      +21     
  Misses         34       34              

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.