Fix NaFlex non-base patch sizes and fail fast on unsupported towers

[chenghuichen]

Summary

Wire OpenCLIP's NaFlex data path to timm's existing patch interpolation support, so --naflex-patch-sizes works with non-base patch sizes.

• Enable timm's parameter-free patch interpolator for compatible NaFlex image towers.
• Pass the model's base patch size and interpolation capability to the data config.
• Flatten base-size patches only; preserve [N, Ph, Pw, C] for non-base train and eval patches.
• Fail early when a tower without interpolation support is used with a non-base patch size.

Problem

OpenCLIP exposes --naflex-patch-sizes, but did not fully connect non-base patches to timm's interpolator. Multi-size
training could reach a model with interpolation disabled, while a single non-base size and eval patches were flattened
before the model could recover their spatial dimensions. These paths failed with:

RuntimeError: mat1 and mat2 shapes cannot be multiplied

Fix

Patch flattening is now based on the model's base patch size rather than the number of configured choices:

base patch      -> [N, P*P*C]
non-base patch  -> [N, Ph, Pw, C]

This rule is applied consistently to training and evaluation. Compatible towers automatically enable timm's existing
interpolator; unsupported towers fail during configuration instead of later in matrix multiplication. The interpolator
adds no checkpoint parameters or buffers.

Tests

Added coverage for 16/32 train and eval paths, forward/backward, checkpoint round trips, automatic enablement, and
unsupported-tower fail-fast behavior.