feat: Add top-k and groupmax activation functions with bias initialization

[psycoplankton]

[FEATURE] Add top-k and groupmax activation functions with bias initialization #13

This PR adds support for top-$k$ and groupmax feature activation methodologies (standard PyTorch autograd versions) as alternative sparsity bottlenecks to the default JumpReLU in CLT-Forge.

Key Changes

1. Configuration (src/clt_forge/config/)

• Added activation_fn (choice of "jumprelu", "topk", or "groupmax") and k fields to both CLTConfig and CLTTrainingRunnerConfig.

2. Model Architecture (src/clt_forge/clt.py)

• Implemented standard autograd-based topk and groupmax activations in the encode method.
• Set requires_grad = False for the log_threshold parameter when using top-$k$ or groupmax activations (since threshold learning is not used).
• Configured the loss method to bypass $L_0$ sparsity penalties and dead feature penalties (setting them to 0.0) under top-$k$ / groupmax configs.
• Robustified _initialize_b_enc by clipping the percentile index with min(..., B - 1) to prevent IndexError on small batches or small feature dimensions.

3. Training & Optimization (src/clt_forge/training/clt_trainer.py)

• Enabled _initialize_b_enc for all training procedures (including top-$k$/groupmax).
• Filtered parameter groups passed to the Adam optimizer to exclude any parameters that do not require gradients (such as log_threshold during top-$k$ runs).

4. Verification & Testing (tests/test_topk.py)

• Created a new unit test suite containing 8 test cases validating:
  • Sparsity guarantees (exactly $k$ active features per token; exactly $1$ per group for groupmax).
  • Correct gradient propagation through encoder/decoder weights.
  • Zeroed out $L_0$ and dead feature loss calculations.
  • Correct behavior of _initialize_b_enc under both topk and groupmax.

---

Verification Results

All new test cases passed successfully:

$ PYTHONPATH=src pytest tests/test_topk.py
============================= test session starts ==============================
collected 8 items

tests/test_topk.py ........                                              [100%]
======================== 8 passed, 4 warnings in 2.97s =========================

# Future Additions
There still needs to be the custom autograd for efficient matrix multiplication as is done in Eleuther's Sparsiy.