Add ResolutionInvariantReadout with FNO integration and gallery examples

[JoshBeacom]

Summary

Adds ResolutionInvariantReadout and an optional readout= parameter on FNO, reconciled with the current main so it merges cleanly.

• ResolutionInvariantReadout maps field outputs to scalar/vector QoIs via spatial reduction (mean/integral) + a linear/MLP projection head (neuralop/models/readouts.py).
• Integrated as an optional readout parameter on FNO, with in_channels↔out_channels validation and a guard against complex_data=True. OTNO explicitly rejects readout (its forward indexes a target mesh and returns (out_channels, n_target), incompatible with (B, C, *spatial) heads).
• Two gallery examples: Darcy scalar QoI and 3-D Poisson–Boltzmann screened energy (demonstrating resolution invariance).

Why this PR (conflict resolution)

The resolution-invariant-readout branch could no longer be merged into main after main gained the GroupNorm (norm_groups) and configurable FNO conv-bias-kernel (conv_bias_kernel) work. This branch now merges main in and resolves the conflict.

The only conflict was in neuralop/models/tests/test_otno.py: both sides appended a new top-level test immediately after test_otno_output_shape, so git could not order them. Resolved by keeping both — they are independent:

• test_otno_rejects_readout (this branch) — OTNO raises ValueError when a readout is passed
• test_otno_group_norm (main) — OTNO honors a custom norm_groups

neuralop/models/fno.py and neuralop/models/otno.py auto-merged cleanly: main's norm_groups/conv_bias_kernel additions and this branch's readout parameter/validation/forward path occupy disjoint regions and all survive intact.

Tests

Run locally (CPU, torch 2.11):

• test_otno.py + test_fno_readout.py — 49 passed
• test_fno.py (incl. main's conv_bias_kernel tests) — 360 passed

(Pre-existing unrelated warnings only: rno.py docstring escape, has_cuda deprecation.)

[JoshBeacom]

Superseded by #7. The conflict resolution was restaged on resolution-invariant-readout-merged for review without touching upstream neuraloperator#719 (which has been restored to its pre-fix state).