Add from scratch pretrain converter

[Cacapice]

Description

Adds a weight converter and loader that bring checkpoints from an external
from-scratch pretraining framework into HookedTransformer, so a
foundation model trained outside the HuggingFace ecosystem can be analysed with
the standard TransformerLens hook API (run_with_cache, resid_post, etc.).

The source architecture is a decoder-only transformer with RoPE (adjacent-pair
rotation), RMSNorm, SwiGLU MLPs, and optional dropless top-k Mixture-of-Experts.
Its checkpoints are not loadable through from_pretrained because they are not
a HuggingFace model; this PR follows the same pattern as the existing
nanogpt / mixtral converters to support them.

Motivation: this is the base-model half of an interpretability project — the
model whose residual stream downstream linear probes read. Loading it into
TransformerLens means the probing side can use the library's tooling directly
instead of a bespoke activation-extraction path.

Fixes # (no issue — new-feature contribution)

Type of change

• New feature (non-breaking change which adds functionality)
• This change requires a documentation update (docstrings included; happy
  to add a docs/ loading example if the maintainers would like one)

What's included

• weight_conversions/maritime_pretrain.py — convert_maritime_pretrain_weights.
  Splits the fused QKV projection into per-head W_Q/W_K/W_V, maps SwiGLU
  gate/up/down onto TransformerLens's W_gate/W_in/W_out, and (for MoE) emits
  the router plus every expert under the names MoE/MoEGatedMLP expect. No
  Q/K reordering is needed because RoPE is applied inside attention, not baked
  into the weights.
• pretrained/maritime_pretrain_loader.py — build_config and
  load_maritime_pretrain. Builds the matching HookedTransformerConfig
  (positional_embedding_type="rotary", rotary_adjacent_pairs=True,
  normalization_type="RMS", final_rms=True, gated_mlp=True, and
  num_experts/experts_per_token/norm_topk_prob for MoE), merges
  tensor-parallel weight shards, and returns a ready HookedTransformer.
• weight_conversions/__init__.py — export the new converter (alphabetical).
• Tests — tests/unit/pretrained_weight_conversions/test_maritime_pretrain.py.

Key correctness detail

The source rotates adjacent RoPE pairs ([x0, x1] -> [-x1, x0]), which maps to
rotary_adjacent_pairs=True (the GPT-NeoX-style d -> (d 2) frequency layout),
not the default half-split convention. Getting this wrong still loads without
error but silently corrupts every activation, so the equivalence tests below are
the real guardrail.

Checklist

• I have commented my code, particularly in hard-to-understand areas
• I have made corresponding changes to the documentation (module + function docstrings)
• My changes generate no new warnings
• I have added tests that prove my feature works
• New and existing unit tests pass locally with my changes
• I have not rewritten tests relating to key interfaces which would affect backward compatibility

Test evidence

test_maritime_pretrain.py asserts bit-exact logit equivalence between the
source model and its converted HookedTransformer, parametrised across the axes
a future change is most likely to break, plus structural, failure-path and
tensor-parallel coverage (13 tests):

• dense and MoE MLPs
• tied and untied embeddings
• four (d_model, n_heads) combinations
• converter key/shape correctness for the MoE branch
• a config/checkpoint shape mismatch raises a clear ValueError (not an opaque
  reshape error)
• tensor-parallel shard merge reconstructs the original weights for
  N ∈ {1, 2, 4} shards

$ python -m pytest tests/unit/pretrained_weight_conversions/test_maritime_pretrain.py -q
13 passed

Observed max|Δlogit| between source and converted model ≈ 1e-5
(floating-point noise) across every parametrisation. Converter validates each
source weight's shape before reshaping, so a mismatched config fails with an
actionable message. Files formatted with the repo's pinned black==23.3.0 /
isort==5.8.0; pylint 10.00/10 and pyright clean on the new modules.

Notes / open questions for reviewers

1. Naming. I've used the project-specific name maritime_pretrain to mirror
   nanogpt. If you'd prefer a generic name (e.g. swiglu_moe_gpt) since the
   converter isn't domain-specific, I'm glad to rename.
2. Loader placement. The loader lives at pretrained/maritime_pretrain_loader.py
   rather than being wired into loading_from_pretrained.py, because these
   checkpoints load from a local run directory, not a hub ID. If you'd rather
   register it as an official model source, point me at the preferred hook.
3. Targeting dev per the contribution guidelines (this is a feature, not a
   fix against the released version).