fix: F8E4M3 CUDA cast kernel name mismatch

DType::F8E4M3.as_str() returned "f8e4m3" but all CUDA kernels use
"f8_e4m3" (with underscore) — e.g., cast_f8_e4m3_bf16. This caused
"named symbol not found" errors for ALL F8E4M3 dtype casts on CUDA.

Fix: return "f8_e4m3" from as_str(), accept both forms in from_str().

Author: jamesbrink

candle-core/src/dtype.rs

@@ -60,7 +60,7 @@ impl std::str::FromStr for DType {
             "f16" => Ok(Self::F16),
             "f32" => Ok(Self::F32),
             "f64" => Ok(Self::F64),
-            "f8e4m3" => Ok(Self::F8E4M3),
+            "f8_e4m3" | "f8e4m3" => Ok(Self::F8E4M3),
             "f6e2m3" => Ok(Self::F6E2M3),
             "f6e3m2" => Ok(Self::F6E3M2),
             "f4" => Ok(Self::F4),
@@ -83,7 +83,7 @@ impl DType {
             Self::F16 => "f16",
             Self::F32 => "f32",
             Self::F64 => "f64",
-            Self::F8E4M3 => "f8e4m3",
+            Self::F8E4M3 => "f8_e4m3",
             Self::F6E2M3 => "f6e2m3",
             Self::F6E3M2 => "f6e3m2",
             Self::F4 => "f4",