### Description
Softmax (formula 1) is like the following:
```math
y_{i} = \frac{exp(x_{i})}{\sum_{i} exp(x_{i})}
```
After applying softmax, each element will be in the range of $(0, 1)$,
and the elements will add up to 1, so that they can be interpreted as
probabilities.
However, in language model, softmax has two issues:
* When all elements are -inf (for example, a whole row is masked when a
query token is padding), the result is not defined since exp(-inf)=0 and
divided-by-zero is encountered in the above formula.
* Why do we need normalize in a way that each query word are treated as
equal important (each row has sum equals to1)?
**Smooth Softmax** (formula 2) is a modified version that introduces a
smooth factor like the following:
```math
s_{i} = \frac{exp(x_{i})}{1+ \sum_{i} exp(x_{i})}
```
This formula could tackle the above two issues:
* It could handle the special case that all elements are -inf: the
result $s_{i}$ is 0 for every element in such case.
* Sum of all elements $\sum_{i}{s_{i}} = \frac{\sum_{i}{exp(x_{i})}}{1+
\sum_{i} exp(x_{i})}$ is in the range of (0, 1), so that we can train
the model to assign different importance to different query words.
Since exponential is prone to overflow or underflow, to get stable
result, formula 3 can be used:
```math
s_{i} = \frac{exp(x_{i} + c)}{exp(c)+ \sum_{i} exp(x_{i} +c)}
```
c can be any value in theory. In practical, choice of constant c shall
avoid $exp(c)$ and $exp(x_{i} +c)$ overflow (or underflow) at the same
time. A reasonable choice is like formula 4:
```math
c=-\max_{i} \{ x_i \}
```
or apply a constraint that c <=0 like the following formula 5:
```math
c=-\max(0, \max_{i} \{ x_i \})
```
The latter one (formula 5) ensures that $s_{i}$ will fallback to formula
2 when all elements are negative.
For CPU provider, smooth softmax is implemented in MLAS. CPU
implementation uses formula 5.
@wangyems implemented the smooth softmax in flash attention for CUDA,
which requires Ampere or newer GPU. The implementation of smooth softmax
in flash attention uses formula 4.
---------
Co-authored-by: Ye Wang
The prefetching instructions (_mm_prefetch) is used to anticipate memory
accesses by prefetching the next row of the input buffer. This
optimization is designed to reduce the impact of memory latency, thereby
enhancing the performance of the MlasComputeSoftmax function. As a
result, the worst-case performance of the OCR model has improved by
approximately 50ms, which equates to a 3% improvement.
Change int32_t->ptrdiff_t when interacting with the threadpool.
Migrate more code from MlasMaskMoveAvx->MlasMaskMoveTableAvx.
Update more code to use FUNCTION_ENTRY macro.
