Disable QDQ related transformers for WebNN EP (#28020)
WebNN EP needs the original DQ/Q nodes preserved. The existing
`session.disable_quant_qdq` option disables QDQ fusion but still allows
constant folding to fold `DequantizeLinear` nodes then losing the
quantization information.
This PR adds a new session option `session.disable_dq_constant_folding`
that prevents `DequantizeLinear` nodes from being individually constant
folded. This option only takes effect when `session.disable_quant_qdq`
is "1"; when QDQ fusion is enabled (default), DQ nodes are already
protected.
This PR also passes `nullptr` for node_unit_map in WebNN EP
`GetCapability` to treat DQ/Q as individual nodes for partitioning. Link
to #27545
