[TIR, TVMScript] Update printer / parser to make T.allocate return buffer var

[vinx13]

• Updated TVMScript syntax of T.allocate to return buffer var.
• Added syntax sugar for T.decl_buffer. When data field is not specified, data will be implicitly created via Allocate stmt.
• Updated the existing test cases. Most test cases can be updated by changing T.allocate to T.decl_buffer. T.allocate in some tests are updated to T.allocate + T.buffer_decl, to maintain the legacy behavior of allocation and implicit buffer declaration (will be followed up in future PR to adopt T.decl_buffer).

Related RFC: apache/tvm-rfcs#87

cc @junrushao1994@Lunderberg@wrongtest-intellif@cyx-6@tqchen

[Lunderberg]

Looks good! I have a couple of questions on the implementation side, but overall it looks very good!

[Lunderberg]

Should we also track which buffers have gone out of scope? If I'm understanding it correctly, a single DeclBufferNode would also allow for usage outside of the DeclBufferNode::body, where I'd expect it to only apply within the scope of the node.

[vinx13]

To answer the question about T.allocate vs T.decl_buffer update:
There are two ways to update the existing TVM scripts.

1. Update T.allocate to T.allocate + T.buffer_decl. This is strictly equivalent to the TIR before.
2. Update T.allocate to T.allocate + T.decl_buffer, (or the syntax sugar form, only a single T.decl_buffer statement without data argument). This adds a DeclBuffer node in TIR, which is not strictly equivalent to the TIR before. This is the preferred way (as it uses explicit DeclBuffer) but might require updating the existing passes. To minimize changes to existing passes in this PR, in the case this method doesn't work out-of-box, we will choose method 1 for now.

[Lunderberg]

That makes sense for minimizing the changes to the generated TIR, so that those can be handled in an independent change. Marking all of those conversations as resolved.

[Lunderberg]

LGTM!