webnn: Use TFLite DEQUANTIZE op when casting float16 to float32

A WebNN graph passing float16 weights to a float32 operator needs to
use a cast() op, however TFLite and its delegates expect that cast to
be done with the DEQUANTIZE op. This change adds a special case for
this scenario. Note, TFLite doesn't support using QUANTIZE for float32
to float16 conversions.

Change-Id: I7aac5e2d594f30433362cfaf0476d3342cec00c9
Reviewed-on: https://chromium-review.googlesource.com/c/chromium/src/+/6557825
Commit-Queue: Reilly Grant <reillyg@chromium.org>
Auto-Submit: Reilly Grant <reillyg@chromium.org>
Reviewed-by: Phillis Tang <phillis@chromium.org>
Cr-Commit-Position: refs/heads/main@{#1461671}