Fix All Reduce not converge on Imagenet dataset

[yiheng]

We find that all-reduce won't converge as fast as one-reduce when train model on imagenet dataset. The root cause is we truncate model parameter from fp32 to fp16 when apply gradient.

This PR fix this issue by keep parameter precision and include some other small changes.

[jason-dai]

We should make broadcastSplit a member variable, which can be used in subsequent sync and sumAndUpdate methods (that is, no need to broadcast it again and again).

[jason-dai]

We should pass the update method to the class constructor instead of the sumAndUpdate method, so that we can broadcast it once (and make broadcastUpdate a class member variable) instead of broadcasting it again and again.

[jason-dai]

@yiheng @dding3 We can simplify the AllReduceParameterManager implementation as follows:

• There is no need to construct the partitionIds table in init() - each task has a unique partition id in the range of 0 to partitionNum-1
• splits can just be an array indexed by the partition id, where each element is just a tuple of (offset, length, blockId). There is no need to compute a map of blockIds for the weight (see below)
• Each task should only write one copy of the slit (for the weight it is responsible for), instead of writing a copy for every other task (as in the current implementation). The blockId for each copy of the split can be precomputed at the driver side.
• Similar, paramTable in sumAndUpdate can be precomputed at the driver side and broadcasted only once

See other comments inline.