flex/moe: UNSLOTH_FLEX_GRAPH_BS env var to override capture bucket ladder

The default ``graph_bs = [1, 2, 4, 8] + range(16, max_bs+1, 16)``
matches the dense FlexInference pattern and is optimal for
power-of-2 batch sizes. For workloads that routinely hit
intermediate sizes (e.g. bs=24, 40, 56), capturing dedicated
buckets improves replay efficiency ~5-10% on those sizes
(they otherwise round up to the next power-of-2 bucket).

Measurements on Qwen3-30B-A3B 4bit, bs=24/40/56 with
compile_walker=True:

| bs | default ladder (round-up)     | UNSLOTH_FLEX_GRAPH_BS=8,16,24,32,40,48,56,64 |
|----|------------------------------:|---------------------------------------------:|
| 24 | (rounds to 32, ~1800 tok/s)   |                                 1836 tok/s  |
| 40 | (rounds to 48, ~2750 tok/s)   |                                 2772 tok/s  |
| 56 | (rounds to 64, ~2700 tok/s)   |                                 3797 tok/s  |

The fine ladder costs a few extra capture calls at startup and
slightly larger CUDA graph pool footprint. Power-of-2 bucket
throughput is marginally lower with the fine ladder (memory pool
is shared across more captures), so it's opt-in via env var.

Invalid values fall back to the default ladder with a log.

unsloth/inference/flex_moe.py

@@ -480,7 +480,22 @@ class FlexMoEInference:
             dtype = self.model.dtype,
             device = self.device,
         )
-        self.graph_bs = [1, 2, 4, 8] + list(range(16, max_bs + 1, 16))
+        # Bucket ladder for CUDA graph capture. Default mirrors the dense
+        # FlexInference pattern. ``UNSLOTH_FLEX_GRAPH_BS=1,8,32,64`` etc
+        # lets you override (useful for tight memory or for bigger batch
+        # bench). Values > max_bs are silently skipped below.
+        _env_bs = os.environ.get("UNSLOTH_FLEX_GRAPH_BS")
+        if _env_bs:
+            try:
+                self.graph_bs = [int(x) for x in _env_bs.split(",") if x.strip()]
+            except ValueError:
+                print(
+                    f"[flex-moe] invalid UNSLOTH_FLEX_GRAPH_BS={_env_bs!r}; "
+                    f"using default bucket ladder"
+                )
+                self.graph_bs = [1, 2, 4, 8] + list(range(16, max_bs + 1, 16))
+        else:
+            self.graph_bs = [1, 2, 4, 8] + list(range(16, max_bs + 1, 16))
         pool = None
         for bs in reversed(self.graph_bs):
             if bs > max_bs: