ensure scalars are float32 in test cases for dropout

[iclementine]

Ensure scalars are float32 in test cases for dropout. Now dropout is computed in the following formula:

out = tensor * mask / (1-p)

In our test cases, the reference is also compute in thay.

But in some cases, 1-p have different byte representations depending on how it is computed.

1. (1 - p) when p is a native python floating point number then cast to float32;
2. (1 - p) when p and 1 are both numpy 32bit floating point number then cast to float32;

Thus, tensor / (1-p) may have have different results when 1-p have different representations.

[tongxin]

LG

[Bowen12992]

LGTM