Need reverse output sequence for backward direction of gru and lstm operator emulations

[BruceDai]

Current lstm operator emulation computes the forward hidden sequence from t = 1 to T, and the backward hidden sequence
from t = T to 1.

While lstm operator of WebNN API Spec says "the third element is the 4-D output tensor of shape [steps, numDirections, batchSize, hiddenSize] containing every output from each time step in the temporal sequence", so we need reverse output sequence for backward direction of lstm operator emulation from [hb_T, ..., hb_1] to [hb_1, ..., hb_T].

I've firstly submitted a PR webmachinelearning/webnn-baseline#123 for preview.

@fdwr @huningxin PTAL, thanks!

/cc @philloooo @fujunwei

[philloooo]

@BruceDai does the gru emulation logic need to be updated too?

      sequence =
        (sequence ? builder.concat([sequence, currentOutput], 0) :
                    currentOutput);

->

      sequence =
        (sequence ? builder.concat([currentOutput, sequence], 0) :
                    currentOutput);

?

[BruceDai]

@BruceDai does the gru emulation logic need to be updated too?

Yes. I'm working to update it. And the solution for gru should be similar to webmachinelearning/webnn-baseline#123.

[inexorabletash]

Now that the PRs linked above have merged, I don't think there is any further work here. Please re-open if I'm incorrect.