Add MaximumMask forcing mask

[ewquon]

Summary

• New MaximumMask(masks...) callable: composes any number of mask objects by
  their pointwise maximum, mask(x, y, z) = max(mask₁(x, y, z), mask₂(x, y, z), …).
• Mask-agnostic — composes GaussianMask, PiecewiseLinearMask, or any other
  callable mask (no hard dependency on CosineRampMask from Add CosineRampMask forcing mask #5574).
• Sits next to GaussianMask / PiecewiseLinearMask in src/Forcings/relaxation.jl,
  exported from Forcings.jl and re-exported from Oceananigans.
• 3-arg (x, y, z) callable matching the existing mask precedent; default Adapt
  covers the struct.
• No new dependencies.

Why

Composing multiple sponge or boundary-nudging zones via MultipleForcings is
additive, which double-strengths the relaxation rate wherever zones overlap —
typically at corners in Davies-style lateral nudging or where an upper sponge
meets a lateral one. MaximumMask saturates at 1 in the overlap region instead,
preserving the user-specified timescale 1/rate throughout the nudging volume.

This is independently useful for users assembling boundary forcing today (no
CosineRampMask required), and pairs cleanly with CosineRampMask (#5574) once
that lands for the canonical Davies recipe.

Tests

• New @testset "MaximumMask construction and pointwise max" in
  test/test_forcings.jl covers:
  • single-mask passthrough
  • pointwise max correctness across a 3D (x, y, z) sweep
  • composition with mixed mask types (GaussianMask + PiecewiseLinearMask)
  • saturation at 1 with overlapping unit-peaked masks
  • ArgumentError on the no-arg constructor
  • summary string format
• New relaxed_time_stepping_max_mask(arch) smoke test runs a Davies-style
  MaximumMask of two GaussianMasks inside a Relaxation on a
  NonhydrostaticModel for one time step, exercising the mask through the
  kernel path; wired into the existing relaxation-forcing testset (CPU + GPU).
• Doctest snippet in docs/src/models/forcing_functions.md showing a
  Davies-style construction; output verified locally.

Test plan

• CPU unit + smoke tests pass locally
• Doctest output verified to match
• CI green (CPU + GPU)
• Docs build green