Simplify Min/Max using ShapeEnv value ranges (#186248)

Summary:
ShapeEnv.simplify previously only had narrow handling for Max(0, x) and
Max(1, x), so value-range facts such as torch._check(u0 <= 5) were not used to
simplify Min(u0, 5). That left guards like Eq(Min(5, u0), u0) data-dependent in
Dynamo even though the ordering was already known.

Generalize the ShapeEnv simplification pass to visit both Min and Max atoms and
remove dominated arguments when ShapeEnv can prove their ordering from value
ranges. The current implementation keeps the reasoning in ShapeEnv, where range
information is available, and uses cheap lower-bound/sign checks before the
heavier bound_sympy path so common Min/Max simplifications do not pay the full
static solver cost. The legacy _maybe_evaluate_static fallback remains limited to
preserving the previous Max(0/1, x) behavior.

This pass runs only in the non-size-oblivious path, preserving the existing
size-oblivious guard behavior. A narrower binary-only change would fix the exact
repro, but using the same proved-ordering rule for n-ary Min/Max avoids leaving
equivalent flattened expressions unsimplified while still requiring ShapeEnv to
prove every removed argument.

Fixes #137096

Generated by my agent

Benchmark Results:

- Mixed cold simplify workload with 400 legacy Max expressions and 400 newly
  handled Min expressions: the earlier solver-heavy PR version was 0.340158s
  best of 5; after the cheap ordering checks and fallback tightening this is
  0.069140s best of 5.
- Unproven Min/Max stress workload with 500 expressions: 0.243360s best of 5
  after fallback tightening, compared with 0.537786s best of 3 before the
  narrowing in that follow-up.
- Legacy Max-only cold simplify workload stayed roughly unchanged in earlier
  benchmarking: main best of 5 was 0.122475s and the patched path was 0.125108s.

X-link: pytorch/pytorch#186248
Approved by: https://github.com/aorenste

Reviewed By: jeanschmidt

Differential Revision: D112034768

fbshipit-source-id: 1578a2df94ca1a242206573f452f2cdcc172e9c6

Author: jansel

userbenchmark/dynamo/dynamobench/huggingface.yaml

@@ -6,6 +6,8 @@ skip:
     - BlenderbotForConditionalGeneration
     - GPTNeoForCausalLM
     - GPTNeoForSequenceClassification
+    # Fails importing transformers.modeling_deberta_v2 with current TorchScript.
+    - DebertaV2ForMaskedLM
     # Fails with even batch size = 1
     - GPTJForCausalLM
     - GPTJForQuestionAnswering