Model optimization: YoloV4

[mywoodstock]

• [ ] change output of all conv2d to ROW_MAJOR -- this will help avoid untilize in halo
• [ ] use block sharding for maxpool2d -- this will enable use of larger #cores
• [ ] upsample is using 1 core, investigate and fix
• [ ] use block sharding for later conv2d ops with larger depth
• [ ] investigate slow unary op
• [ ] minimize i2s, s2i, tilize, untilize, with/without padding
• [ ] use generalized concat to avoid interleaved (blocked by unsupported concat)
• [ ] enable optimization flags in conv2d: double buffering, split reader
• [ ] use bfloat8_b where possible
• [ ] switch to LoFi fidelity

[shwetankTT]

https://docs.google.com/spreadsheets/d/1DkhdEmNB8ZVjb0LbgxGXhtoyRxFtySBl1jBGwheEs_U/edit?usp=sharing

Tried multiple permutations and combinations for using row_major_layout the current config is giving us the best output. This is for Down1 (submodel of yolo4)

Concern: Unary Ops taking too much time.

[shwetankTT]

Neck perf improvement. https://docs.google.com/spreadsheets/d/1Zhm2PSm-GVqM5ujH3fru5EwEWG1vN6eqF5kA0y_AxOE/edit?usp=sharing