Added pattern rewrite to optimize float ops

[LeslieXMOS]

Pattern that optimizes float ops with dequantize/quantize ops.

Optimization includes

• MoveDequantForwardAndReplaceSqrtPattern: move dequantize op forward over sqrt pattern and replace the sprt pattern with a lookup op
• MoveDequantForwardOverUnaryOpPattern: move dequantize op forward over unary op (Abs, Neg, Log, Logistic, SumOp), readjust zero point and scale when over AbsOp
• MoveDequantForwardOverSameInputOpPattern: move dequantize op forward over op where lhs and rhs connected to same tensor, readjust zero point and scale for MulOp output
• FoldDequantMulQuantOpPattern: fold dequantize -> mul (rank0 const) -> quantize with a single quant, the output scale is adjust with the mul const multiplier
• FoldDequantDivQuantOpPattern: remove lhs and rhs dequantize ops of div op if it's output is also connected with a quant op, rewrite this pattern with lhs directly connect to mul op and rhs connected to a lookup to perform 1/rhs and connected to mul op, the mul output replace the quant op output
• FoldDequantQuantPairPattern: fold dequantize -> quantize pair into a single quantize op