Refactor the original "PassManager" class, and derive the "GraphKernelPassManager"
GraphKernel's ir files are dumped into a new sub-directory "graph_kernel" in the original "verbose_ir_files"

All GraphKernel's passes are divided into 3 levels, and controlled by the flag "opt_level" by default.
when the opt_level is greaterequal to the pass's level, this pass will run.
The default "opt_level" is 2 when GraphKernel is enabled.
Levels:
  1. Basic features, like cluster, splitter, and some preprocess, postprocess.
  2. All stable features, mainly includes the optimization passes.
  3. Experimental features, like stitch-fusion, parallel-fusion.

The two flags "enable_pass" and "disable_pass" are available in this commit.
User can manually enable some passes when it's disabled by "opt_level", or disable the enabled passes,
by specifying that pass in this format: "stage_id.pass_id" or "stage_name.pass_name", multiple passes are separated by comma(",")
the stage/pass index and stage/pass name can be found from the ir filename.
e.g.   "--enable_pass=cluster.graph_kernel_expander,1.1,1.2"

Others:
  1. the pass "tensor_promotion" is not useful, remove it.
  2. put the pass "InsertPadOps" before "ArithmeticSimplify".

From: @lingyunli63
Reviewed-by: 
Signed-off-by:

1. clone PrimitivePtr before setting attr, operators may use the same shared_ptr.
2. bugfix for the reduce-reduce pattern when keep_dims is true.

From: @looop5
Reviewed-by: @anyrenwei,@gaoxiong1
Signed-off-by: @gaoxiong1

From: @looop5
Reviewed-by: @ckey_dou,@gaoxiong1
Signed-off-by: @gaoxiong1

From: @dayschan
Reviewed-by: @gaoxiong1,@anyrenwei
Signed-off-by: @anyrenwei

From: @liu-yihong536
Reviewed-by: @gaoxiong1,@anyrenwei
Signed-off-by: @anyrenwei

In some place, we directly visit the node's input kernel_info.
but when the node's input is a Parameter, it will fail.

From: @chenlei_autodiff
Reviewed-by: @gaoxiong1,@ckey_dou,@gaoxiong1,@ckey_dou
Signed-off-by: @ckey_dou,@ckey_dou

From: @dayschan
Reviewed-by: @gaoxiong1,@ckey_dou
Signed-off-by: @ckey_dou

From: @coding2020
Reviewed-by: @gaoxiong1,@ckey_dou
Signed-off-by: @ckey_dou

From: @wenfangpei
Reviewed-by: @gaoxiong1,@ckey_dou
Signed-off-by: @ckey_dou

From: @dayschan
Reviewed-by: @ckey_dou,@gaoxiong1
Signed-off-by: @gaoxiong1

From: @dayschan
Reviewed-by: @ckey_dou,@gaoxiong1
Signed-off-by: @gaoxiong1

When creating the new GKDropout, the original process used the Dropout's input's kernel_info to infer the kernel_info of new node.
but if the input is a Tensor, then it hasn't kernel_info, which causes the GKDropout's input_kernel_info to be invalid.
In this submit, I manually set the kernel_infos to ensure they are always valid.

From: @dayschan
Reviewed-by: @ckey_dou,@gaoxiong1
Signed-off-by: @gaoxiong1

From: @jiaoy1224
Reviewed-by: @gaoxiong1,@gaoxiong1,@ckey_dou
Signed-off-by: @ckey_dou

the original pass used mng->RemoveRoot and mng->KeepRoot to maintain manager,
but in some network it works slowly. so we use mng->Replace instead.

From: @wenfangpei
Reviewed-by: @gaoxiong1,@ckey_dou
Signed-off-by: @ckey_dou

From: @wenfangpei
Reviewed-by: @gaoxiong1,@ckey_dou,@gaoxiong1,@ckey_dou
Signed-off-by: @ckey_dou,@ckey_dou

From: @tronzhang
Reviewed-by: @gaoxiong1,@ckey_dou
Signed-off-by: @ckey_dou

Use a disjoint-set to maintain the clusters before building anf-graph, to speed up the process of building graph.
Dump the change of nodes to file if the graphkernel flag "dump_as_text" is set.

The algorithm of CheckCircle is unchanged.
Reuse the FuseNodesToSubGraph interface.

From: @wenfangpei
Reviewed-by: @ckey_dou,@gaoxiong1
Signed-off-by: @ckey_dou

From: @lingyunli63
Reviewed-by: @ckey_dou,@gaoxiong1
Signed-off-by: @gaoxiong1

From: @r1chardf1d0
Reviewed-by: @gaoxiong1,@ckey_dou
Signed-off-by: @ckey_dou

From: @wenfangpei
Reviewed-by: @gaoxiong1,@ckey_dou
Signed-off-by: @ckey_dou

From: @zengzitao
Reviewed-by: @ckey_dou,@gaoxiong1
Signed-off-by: @gaoxiong1

fix compiling

From: @hwhewei
Reviewed-by: @zh_qh,@ginfung
Signed-off-by: @zh_qh