Enable custom op and avoid graph breaks

[ZhangLirong-amd]

No description provided.

[ZhangLirong-amd]

custom op define only support types like Tensor, Tensor[], int[], int, float, bool... Not support like Enum, Object and so on..
The list of op custom op not support and we need to change:

register_graph_buffers  : not support list[str], list[list[int] as input schem only remain this one 

finished op:
fmoe_int8_g1u0,  
fmoe_g1u1 , 
fmoe_g1u1_tkw1, 
fmoe_fp8_blockscale_g1u1,
moe_stage1_g1u1
mha_varlen_bwd : dict
hipb_mm
hipb_findallsols

For fake function, we can not infer the shape in @compile_ops, thus we return None for all op which is return Tensor before.
The following op list is we need to modify it's api to return None.
Part of them is finished, the remaining list:

add
sub,
div,
sigmoid : done
tanh : done
all_reduce_asm_
all_reduce_rmsnorm_
all_reduce_rmsnorm_quant_
get_graph_buffer_ipc_meta
allocate_meta_buffer
get_meta_buffer_ipc_handle
module_mha_varlen_fwd
hipb_mm
rocb_mm
mha_varlen_fwd  :list
pa_fwd_naive :done
pa_fwd_asm  :done
module_moe_asm

But these maybe not all problematic ops. There may be more during the testing process.

update some new problematic ops need return None:

mha_batch_prefill  :list
fmha_v3_fwd: list
fmha_v3_bwd
mha_bwd
mha_fwd

[ZhangLirong-amd]

problem in accuracy:

add
mul
sub
div
get_graph_buffer_ipc_meta
get_meta_buffer_ipc_handle
aiter_sigmoid
aiter_tanh
pa_fwd_naive
pa_fwd
mha_batch_prefill
mha_bwd
mha_fwd
mha_varlen_bwd
mha_varlen_fwd
all_reduce_asm
all_reduce_rmsnorm
all_reduce_rmsnorm_quant
fmha_v3_bwd
fmha_v3_fwd
fmha_v3_varlen_bwd

RocSolIdxBlas： rocb_mm

module_moe_asm

hipb_mm