Merge pull request !1280 from liubuyu/master

Merge pull request !1279 from gongchen/fix_custom_op_case

Merge pull request !1058 from sunsuodong/lstm_ops

be optimized as it will generate different values each time.

Merge pull request !1015 from candanzg/me_with_shape

Signed-off-by: candanzg <zhangshucheng@huawei.com>

Merge pull request !1091 from chenweifeng/cast

Merge pull request !1082 from hewei/support_cont_break

To handle 'break' and 'continue' statement, a loop context is pushed
to a stack before we parse the loop body, and pop it after body parsed.
When a 'break', 'continue' statement is encountered, we retrieve current
loop contex from the stack, and let the current block jump to the end
block or header block;

For 'break' statement, we added an extra 'end_block' follow the 'after_block',
because 'after_block' is called from a ContionalJump in 'header_block', it can
not be set as jump target from other place. to support 'break', we let loop
body jump to the 'end_block' at the 'break' point. and 'after_block'
maybe a good place to handle loop 'else' clause in the future.

Handle 'pass' is simple, just bypass it when doing parse.

Merge pull request !1057 from chenweifeng/matmul

Merge pull request !1040 from zhangdengcheng/master

Merge pull request !370 from chenweifeng/unsorted_segment_sum

Merge pull request !994 from chenweifeng/broadcast_unequal_dims

Merge pull request !1021 from chenweifeng/broadcast_grad

Merge pull request !948 from chenweifeng/logsoftmax

Merge pull request !920 from wsc/lossscale_script

Merge pull request !322 from chenweifeng/rmsprop

Merge pull request !323 from chenweifeng/concat

Merge pull request !324 from chenweifeng/slice

Merge pull request !372 from chenweifeng/batchmatmul

Merge pull request !502 from chenweifeng/gelu

Merge pull request !516 from chenweifeng/tanh

Merge pull request !945 from chenweifeng/unary

Merge pull request !812 from jinyaohui/train_eval