[autoparallel] Draft for mix gather

[zhang677]

What's new? Add a one-step transformation called mix-gather for:

Src	Dst
S0S1	RR
S1S0	RR
S01R	RR
RS01	RR

Why do we need this? Reduce the communication cost. Assume $\beta_1 \gt \beta_0$, $M$ is the communication size. Cost for S0S1=>S0R=>RR is $\frac{M}{n_1n_0}\times\frac{n_1-1}{n_1}\times\beta_1 + \frac{M}{n_1}\times\frac{n_0-1}{n_0}\times\beta_0$ Cost for S0S1=>RR is $\frac{M}{n_0n_1}\times\frac{n_0n_1-1}{n_0n_1}\times\beta_1$

Pitfalls Peak memory increases by the size of a tensor for S0S1=>RR and S1S0=>RR

[YuliangLiu0306]

Could you please push the unit test file for this feature?