Allow non-standard output shape

When the final output is a non-standard shape, the compile passes will add a contiguous_kernel to make it into a standard shape.

Following python script can reproduce this issue:

import migraphx

p = migraphx.program()
mm = p.get_main_module()
mgx_shape = migraphx.shape(lens=[2, 1, 4], type='float_type')
inp = mm.add_parameter('x', mgx_shape)
out = mm.add_instruction(migraphx.op('multibroadcast', out_lens=[2, 4, 4]),
                         [inp])
mm.add_return([out])

print("Uncompiled Program: ")
print(p)

p.compile(migraphx.get_target('gpu'), offload_copy=False)

print("Compiled Program: ")
print(p)

Output:

Uncompiled Program: 
module: "main"
x = @param:x -> float_type, {2, 1, 4}, {4, 4, 1}
main:@1 = multibroadcast[out_lens={2, 4, 4}](x) -> float_type, {2, 4, 4}, {4, 0, 1}
main:@2 = @return(main:@1)


Compiled Program: 
module: "main"
main:@0 = check_context::migraphx::version_1::gpu::context -> float_type, {}, {}
main:@1 = hip::hip_allocate_memory[shape=float_type, {0}, {1},id=main:scratch] -> float_type, {0}, {1}
main:#output_0 = @param:main:#output_0 -> float_type, {2, 4, 4}, {16, 4, 1}
x = @param:x -> float_type, {2, 1, 4}, {4, 4, 1}
main:@4 = multibroadcast[out_lens={2, 4, 4}](x) -> float_type, {2, 4, 4}, {4, 0, 1}
main:@5 = gpu::code_object[code_object=13640,symbol_name=contiguous_kernel,global=1024,local=1024,](main:@4,main:#output_0) -> float_type, {2, 4, 4}, {16, 4, 1}
main:@6 = @return(main:@5)

When integrating with PyTorch, some pipelines require the accelerator to output non-standard shapes. In the able example the compiled program should output float_type, {2, 4, 4}, {4, 0, 1} to be compatible.

Start by running the Python reproducer with the GPU target and compare the uncompiled and compiled shapes. Trace the compile passes that introduce contiguous_kernel and the output parameter; the fix is done when the compiled program preserves the non-standard output shape {2, 4, 4}, {4, 0, 1} and remains compatible with the reported PyTorch pipeline.