[TIR][Hybrid] fold attr&allocate/realize

[spectrometerHBH]

see also #95

cc @tqchen @junrushao1994 @Hzfengsy

[tqchen]

feedback: use allocate overload

[spectrometerHBH]

I realize a problem when trying overloading, these two different tir.allocate follow different print formats.

# folded allocate
a = tir.allocate("float32", [1024], "global")
with tir.allocate("float32", [1024], "global") as a:
# pure allocate
tir.allocate(a, "float32", [1024])
with tir.allocate(a, "float32", [1024]):

And now the registration mechanism will discriminate these two formats, not allowing one registered function to follow two different sets of formats, and I don't want to break this limit here. Meanwhile, they accept different arguments.

Since tir.allocate() will always have a corresponding tir.attr() above, we may just remove pure tir.allocate() in Hybrid Script.

[tqchen]

@spectrometerHBH your change sounds good to me. We might want to think about the argument location of tir.allocate and tir.match_buffer. To make things consistent, we might want to move the dtype after the extents

[spectrometerHBH]

@tqchen I see. And it seems that tir.realize is also foldable with tir.attr(..., "realize_scope"). We may also want to fold them together.

[spectrometerHBH]

cc @tqchen @junrushao1994 @Hzfengsy

[junrushao]

Will review today!