Implement the MODEL emg2pose

We need to think of a nice way to implement the visualization of the data, in the case of pose, but this is another issue.

Ideally, it would be some repr with HTML to be able to see only in Jupyter Notebook, like the eegdash viewer.

This is important for the competition.

Paper/protocol review for this implementation request:

Primary paper: emg2pose, arXiv:2412.02725 / NeurIPS 2024.
Future qualification target: Table 5, held-out-user tracking:
7.7 +/- 1.0 degrees and 10.3 +/- 1.5 mm.
Evaluation unit: aggregate the 5-second trajectory errors per user, then
report mean +/- SD across the 20 held-out-user test users.
Population and split: the corpus has 193 users; the official held-out-user
split is 158 train, 15 validation, and 20 test users. Any future comparison
must retain that split rather than replacing it with random window splitting.
This cycle is protocol-only: no experimental reproduction, checkpoint
qualification, or result claim will be made.
The public corpus is approximately 431 GiB, so it will not be copied into
/home/tau/baristim/iceberg during this cycle.
The official emg2pose repository is CC BY-NC-SA 4.0 and the UmeTrack
dependency is also noncommercial. Do not copy source, rehost weights, or
vendor data into the BSD-3 Braindecode package without an explicit
maintainer/legal clean-room and distribution decision.
Visualization remains outside this issue and should be proposed separately
only after the model/API boundary is accepted.

Before implementation starts, this issue therefore needs:

Navigation Menu