[ChrisW09]

Fixes #555

Problem

The MambaTab pages still described the behaviour of every release before 2.1.0, where MambaTab.forward skipped its Mamba block. 83508ee ("fix(architectures): apply the Mamba block in MambaTab.forward", released in 2.1.0) wired the block in, but the docs kept saying:

• it is "not called in the current forward";
• the Mamba settings "do not affect the current forward pass";
• mamba_version "has no effect in the current release";
• MambaTab is only a "projected-feature baseline", and Mambular is the choice "where the Mamba block must be active";
• axis is a "Temporary unsqueeze axis before normalization".

Change

Docs only:

• docs/model_zoo/stable/mambatab.md: the overview, forward steps, diagram, building-blocks table and implementation notes now include the Mamba block and explain the length-one-sequence (axis=1) vs. batch-as-sequence (axis=0) layout. A short note says that releases before 2.1.0 (and the earlier mambular package) skipped the block, so older MambaTab results describe a projected-feature network.
  • The "Key settings" table adds n_layers, d_state and expand_factor, and explains that axis=0 scans the rows of a batch as one sequence, so a row's prediction depends on the rows before it in the batch.
  • "When To Use" recommends MambaTab as a compact Mamba baseline over one projected token per row, and Mambular or MambAttention for sequence modeling over per-feature tokens.
• docs/model_zoo/stable/index.md: the recommendation describes MambaTab as a compact Mamba baseline over one projected token per row.

Verification

A forward hook on MambaTab.mamba fires with input shape (batch, 1, d_model) on main. The defaults in the table match MambaTabConfig (n_layers=1, d_state=128, expand_factor=2), and v2.0.0, v1.8.0 and mambular v1.3.2-v1.5.0 never call self.mamba in forward.

Checks

• just test: 3236 passed, 46 skipped, 28 xfailed; coverage 91.29%
• just docs: build succeeded (warnings treated as errors)
• just check: all hooks passed (ruff, prettier, pyright)