feat(training): expose map_fusion_mode as a train_il parameter

[ChristophBank]

Problem

AutoE2E.__init__ has accepted map_fusion_mode since #94, and
MAP_FUSION_REGISTRY offers residual / cross_attn / deformable. But
train_il never forwarded it, so every run trained the constructor default —
residual — regardless of intent, and nothing in the run log said so.

That makes the "Map Feature Fusion: Residual, Attention" axis of #168
unreachable through the sanctioned path. A contributor comparing the two either
compared residual against residual, or left train_il for a custom script and
lost comparability with everyone else. It is a silent failure: the run looks
healthy and produces a plausible number.

What this adds

• MapFusion enum beside the existing Backbone enum
• map_fusion_mode on train_il and wf_train_il
• the value in checkpoint_config, so evaluation rebuilds the same
  architecture — _model_kwargs feeds that dict into AutoE2E(**config), so
  without the key a non-default run is reconstructed with the constructor default
  and loads mismatched weights
• model/map_fusion_mode to MLflow, so runs are distinguishable after the fact

The default reproduces the previous behaviour exactly. A run that does not pass
the argument is unchanged.

Deliberately not included

planner_mode is left to #172, which already adds it to train_il along
with the compute_planner_loss wiring. Only the map-fusion half is unique to
this PR, so it stays narrow rather than fixing planner_mode as well.

BEV grid size is left to #188 (camera_bev_size), which touches the same
train_il signature block. This PR adds one parameter next to backbone and
should rebase cleanly either way; happy to rebase behind whichever lands first.

Tests

Model/tests/test_map_fusion_selection.py — a new file rather than an addition
to test_workflow_training_lifecycle.py, which both #172 and #188 modify. Could also integrate into test_workflow_training_lifecycle.py if prefered.

Covers: the pass-through into AutoE2E, the workflow wiring, the checkpoint
round-trip through _model_kwargs, and that every enum value is a registry key.
Verified to fail without the change (5 of 8 fail).

Note on CI: Model/tests currently has 40 pre-existing failures on main
(reasoning, planner, projection, auto_e2e), unrelated to this PR — the counts
are identical with and without it. Nothing in Platform/pipelines fails.

cross_attn is a trap at the contract resolution

Worth knowing before anyone selects it: MapCrossAttentionFusion.forward raises
above 4096 BEV tokens, and the KITScenes contract grid is 256×256 = 65,536. So
deformable is the attention-based option at that resolution. The enum exposes
cross_attn anyway because it is valid at smaller grids, and the guard fails
loudly rather than silently.

[riita10069]

Thanks for the fix, this was a clear bug. Great.