Fix b_dec being double-counted in the TopK AuxK loss

[robbiebusinessacc]

Fixes #132.

The AuxK loss (dead-latent revival, Appendix B.1) is computed against the
residual target e = y - sae_out, where sae_out already includes the
decoder bias b_dec. The second decoder pass that produces e_hat, however,
called self.decode(...), and decode() adds b_dec (return y + self.b_dec). This double-counts b_dec: e_hat ends up shifted by an extra
b_dec, so dead latents are pulled toward e - b_dec rather than e, and an
unintended gradient is placed on b_dec.

This is the normal case in practice, since b_dec is initialized to the data
mean (nonzero), so every TopK SAE/transcoder trained with auxk_alpha > 0 is
affected.

Fix

Compute e_hat by calling decoder_impl directly with W_dec — e_hat = decoder_impl(auxk_indices, auxk_acts.to(self.dtype), self.W_dec.mT) —
skipping the + b_dec that decode() adds. When b_dec == 0 the result is
unchanged; when b_dec != 0 the AuxK target is now correct.

Testing

Added tests/test_auxk_loss.py, a CPU-only test (uses the eager decoder
fallback, no GPU required) that:

• builds a SparseCoder with a nonzero b_dec,
• runs forward(x, dead_mask=all-True),
• asserts out.auxk_loss equals the manually decoded target without re-adding
  b_dec, and
• asserts the previous (buggy) + b_dec formulation produces a different
  value.

The test fails against the previous code (Expected 2.0363 but got 2.7278)
and passes with the fix. ruff check, ruff format --check, and black --check are all clean on the changed files.

[CLAassistant]

All committers have signed the CLA.