[Not for merge] Add Repeat-K mechanism to Zipformer

[yfyeung]

Keeping a separate vector as an nn.Parameter in the model, and adding it times k/(1+k) to the embedding before the ReLU during training and decoding.

It only supports greedy_search and greedy_search_batch.

Result on LibriSpeech 100h

pruned_transducer_stateless7 -> pruned_transducer_stateless9

test-clean	test-other	config
6.48 -> 6.64	17.03 -> 17.50	epoch 30 avg 1
6.43 -> 6.49	16.65 -> 16.85	epoch 30 avg 2
6.31 -> 6.32	16.57 -> 16.53	epoch 30 avg 3
6.31 -> 6.28	16.39 -> 16.23	epoch 30 avg 4
6.26 -> 6.23	16.22 -> 16.01	epoch 30 avg 5
6.22 -> 6.18	16.13 -> 15.89	epoch 30 avg 6
6.24 -> 6.15	16.03 -> 15.80	epoch 30 avg 7
6.17 -> 6.13	15.96 -> 15.78	epoch 30 avg 8
6.11 -> 6.07	15.91 -> 15.79	epoch 30 avg 9
6.10 -> 6.03	15.88 -> 15.79	epoch 30 avg 10
6.10 -> 6.03	15.81 -> 15.73	epoch 30 avg 11
6.06 -> 6.02	15.77 -> 15.78	epoch 30 avg 12
6.08 -> 6.04	15.77 -> 15.81	epoch 30 avg 13
6.05 -> 6.05	15.79 -> 15.83	epoch 30 avg 14
6.02 -> 6.07	15.81 -> 15.80	epoch 30 avg 15

pruned_transducer_stateless9

test-clean & test-other	sum	config
6.03 & 15.73	21.76	epoch 30 avg 11
6.02 & 15.78	21.80	epoch 30 avg 12
6.03 & 15.79	21.82	epoch 30 avg 10
6.04 & 15.81	21.85	epoch 30 avg 13
6.07 & 15.79	21.86	epoch 30 avg 9
6.07 & 15.8	21.87	epoch 30 avg 15
6.05 & 15.83	21.88	epoch 30 avg 14
6.06 & 15.82	21.88	epoch 30 avg 16
6.13 & 15.78	21.91	epoch 30 avg 8
6.05 & 15.86	21.91	epoch 30 avg 17
6.05 & 15.87	21.92	epoch 30 avg 18
6.15 & 15.8	21.95	epoch 30 avg 7
6.1 & 15.94	22.04	epoch 30 avg 19
6.18 & 15.89	22.07	epoch 30 avg 6
6.17 & 16.01	22.18	epoch 30 avg 20
6.23 & 16.01	22.24	epoch 30 avg 5
6.28 & 16.23	22.51	epoch 30 avg 4
6.32 & 16.53	22.85	epoch 30 avg 3
6.49 & 16.85	23.34	epoch 30 avg 2
6.64 & 17.5	24.14	epoch 30 avg 1