Remove constraint on max sequence length

[yawenzzzz]

When rslearn aligns timestamps across multiple modalities, the resulting temporal sequence can be > 12 timesteps. Previously, CompositeEncodings pre-allocated a fixed-size position encoding table based on max_sequence_length (default 12), which will raise an error when the input had more timesteps than the table size. The fix here computes the temporal position encoding on-the-fly from the actual number of input timesteps, removing the fixed limit. It's also backward compatible since we directly drop the pos_embed parameter from old checkpoints.

[favyen2]

thanks for updating this!

[robmarkcole]

I'm quite keen to test this, @yawenzzzz do you have any results to share?