[dataset] support bucket by seq length

[Mddct]

follow this pr #2316

TODO:

• unit test
• aishell training

[Mddct]

it works

    batch_conf:
        batch_type: 'bucket' # static or bucket or dynamic
        bucket_boundaries:  [500, 1000, 1500]
        bucket_batch_sizes: [82, 64, 32, 16]

[500 1000 15000] 代表：
[0,500) [500-1000) [1500, sys.max]
-> 5秒的一个桶 （出来数据为batch size=82）
-> 5-10 的一个桶
-> 10-15 的一个桶
-> 15- 之外的一个桶
这些桶分别产生对应 batch_size的数据，

train_transformer.yaml

以前一个epoch需要1100个batch

现在：

raw 模式训练时间： 14h13min

结果和这里一致： #2316 (comment)

[Mddct]

以下增大bucket的测试：

    batch_conf:
        batch_type: 'bucket' # static or bucket or dynamic
        bucket_boundaries:  [500, 1000, 1500]
        bucket_batch_sizes: [128, 64, 32, 16]

• raw模式， 训练时见14h11min

• shard 模式， 训练时间13h10min

[Mddct]

这里列下上边两个表格中提到的一些指标， 直观一些

batch size	data type	训练时间	att/rescore/ctc greedy/ctc beam wer
old io static 26	raw	20h 5min	5.64/5.31/5.88/5.88
static 26	raw	17h20min	5.56/5.25/5.89/5.89
static26	shard	14h25min	5.53/5.27/5.86/5.86
bucket_boundaries: [500, 1000, 1500] bucket_batch_sizes: [82, 64, 32, 16]	raw	14h13min	5.58/5.27/5.82/5.82
bucket_boundaries: [500, 1000, 1500] bucket_batch_sizes: [128, 64, 32, 16]	raw	14h11min	5.53/5.24/5.85/5.85
bucket_boundaries: [500, 1000, 1500] bucket_batch_sizes: [128, 64, 32, 16]	shard	13h 10min	5.59/5.20/5.73/5.72

[robin1001]

Great job!