[AMD] revamp range analysis

[yangshuxin]

This change fix bugs in range-analysis, and let buffer-ops use the range-analysis result to decide if it's legal to convert memory-op to buffer-ops.

It fixes https://github.com/ROCm/triton-internal/issues/1180 and ROCm#871.

The highlight are following:

• Range Analysis

  • fix the way to use tl.assume. Previously, it does not consider the control flow relationship between, say tl.assume x > 0 and the location of occurrence of x.
  • correct the value range of make_range(begin, end), previous vr is [begin, end], now is [begin, end-1]. Small change in concept incur huge change the regression test.
  • being able to work for some SCP operations, providing a way to side-step MLIR's frameworks' limitation (it does not understand SCF).
  • add a tech note to the file illustrate how to side-step the dataflow frameworks' limitation.

• Buffer-ops

• for large tensor (>2G), remove the ad-hoc, and mistaken range-analysis in the pass. It only relies on the result of the range-analysis pass.
• previous, buffer-ops pass only check element-index > 0. The right condition is byte-offset in [0, 2G-element-size].
• Previous there is a similar work here [AMD] Update BufferOps Non-Negative Check to be in Bytes #7908, contributed by @njriasan . My change to this part is similar but fix some bugs in PR7908 (.e.g. lattice could be nullptr), and update large number of testings. That being said, now that @njriasan made the first change, credit for the part belong to him.

[yangshuxin]

I set change the status to "ready for review" despite failures on Nvidia platform. My change is only about AMD platform, I guess my change is not the culprit.