feat(datapipes): add ComputeUnitGlobalVector transform

[peterdsharpe]

What

A small opt-in transform, ComputeUnitGlobalVector:

• Computes the unit direction v / |v| of a vector stored in global_data.
• Stores it under a new key; the source vector is not modified.
• Works on Mesh and, at the domain level, on DomainMesh.
• Rejects bad input loudly: missing field, zero or non-finite vector, integer dtype, batched (B, 3) input, and output_field == vector_field all raise clear errors.

Why

• Pipelines often nondimensionalize fields and geometry but leave global_data in physical units, because the physical values are needed later to convert predictions back (e.g. force coefficients).
• A model that takes a freestream velocity as input then sees a large dimensional vector (tens of m/s), which can destabilize scale-sensitive models at initialization.
• The unit direction is the scale-free input such models should see. Placed before a rotation augmentation with transform_global_data: true, it rotates together with the source vector and the geometry.

Why a draft

This is new API rather than a bug fix (the bug fix it was originally bundled with is #1820), so a few design questions are worth settling first:

• On DomainMesh, should the direction also be broadcast to sub-mesh global_data (as SetGlobalField does), or stay domain-level only as implemented?
• For datasets that vary inflow speed, should this ship with a companion transform for a nondimensional magnitude (e.g. |v| / v_ref), so speed information is not silently dropped?
• Should batched (B, 3) vectors be normalized per row? (Currently rejected with an error; transforms run per-sample, where vectors are (3,).)
• Does this belong in core, or should it stay recipe-local until there is a second consumer?

Tests

New test/datapipes/transforms/test_compute_unit_global_vector.py (10 tests):

• direction math on Mesh and DomainMesh
• source vector and sub-mesh global_data unchanged
• output keeps the source dtype, so a later rotation composes without mixed-dtype errors
• clear errors on missing field, zero vector, integer dtype, batched input, and same input/output field
• resolvable via the ${dp:...} config resolver

Full transforms suite passes (287); ruff check/format clean.

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.