RNNprofiler: fix gates size retrieval logic in _rnn_flops

[pinstripe-potoroo]

Hello,

Sorry, this MR fixes a bug I introduced myself by mistaking the shape of w_ih to be (input_size, hidden_size), when in reality it is (N*hidden_size, input_size), where N differs depending on the RNN cell type (1 if RNN, 3 if GRU, 4 if LSTM).

To avoid complexifying the logic, I went back to something very close to original implementation, simply multiplying by 2 for FLOP (instead of MAC) and subtracting what I called gates_size = N * hidden_size to avoid counting the bias twice.

Thanks, and sorry for introducing another bug through my previous fix...

Cheers

[pinstripe-potoroo]

Small bump on this MR as the FLOP count is quite off for RNN (because of my previous MR). Although I'm probably the only one still using RNNs these days 😅 I know the MR queue is huge, so no pressure.

[pinstripe-potoroo]

is there anything I can do to help? would you want me to add a unit test?

[pinstripe-potoroo]

small ping on the MR, glad to help if anything is blocking the approval / merging

[pinstripe-potoroo]

thank you @tjruwase and sorry again for introducing a bug!