Pass normalized speed mode to fp16 accumulation gate

The raw speed_mode string was forwarded to _enable_fp16_accumulation, so a
case-variant like MAX failed the speed_mode != SPEED_MAX check and wrongly
disabled fp16 accumulation on float16 pipelines. Forward the normalized mode
and cover the case-insensitive path in the test.

studio/backend/core/inference/diffusion_speed.py

@@ -218,7 +218,7 @@ def apply_speed_optims(
     # and the UNSLOTH_DISABLE_FP16_ACCUM kill switch.
     if on_cuda:
         applied["fp16_accum"] = _enable_fp16_accumulation(
-            family, logger, dtype = getattr(target, "dtype", None), speed_mode = speed_mode
+            family, logger, dtype = getattr(target, "dtype", None), speed_mode = mode
         )
 
     # --- the compile lever, remapped per tier ----------------------------------------

studio/backend/tests/test_diffusion_speed.py

@@ -485,7 +485,7 @@ def test_fp16_accum_allowed_on_fp16_dtype_under_max(monkeypatch):
         _target(dtype = "float16"),
         is_gguf = True,
         family = _family(),
-        speed_mode = "max",
+        speed_mode = "MAX",
     )
     assert applied["fp16_accum"] is True
     assert torch.backends.cuda.matmul.allow_fp16_accumulation is True