perf: switch to FP16 weights — halve download sizes with zero accuracy loss

Candle auto-converts FP16 to FP32 at load time (via absorb_bn).
Verified: max_diff=0.0 across all 5 scales.

| Model | FP32 | FP16 |
|-------|------|------|
| n | 10 MB | 5 MB |
| s | 40 MB | 20 MB |
| m | 88 MB | 44 MB |
| l | 106 MB | 53 MB |
| x | 237 MB | 118 MB |

weights/yolo26l.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:d811fe39d35e9ad04b6255c986dbac2affe0f22c1d600840c40776906f8b1d8f
-size 105590696
+oid sha256:d62b138c6fd81c9dc5dfc3dd1bd98798cd5b07d7586840020a3cf25dc6a4ba6c
+size 52849388

weights/yolo26m.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:f2496ab1c02e582bbebcff61b9dd890dae031f2b5afa44c178217d66a0a7c671
-size 87886672
+oid sha256:b7f99e207d37759ffcdfe15ab2b557d70bc1dbd996e709b095e54a12b6442f1e
+size 43981312

weights/yolo26n.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:c63863f7ad354234e13422d60642dd3344962543eba50868bd6847a15c5433c2
-size 10436600
+oid sha256:a856e63cae6f6c79b509867920172f9becc5fb537acd39aef7b5151638242e2c
+size 5252420

weights/yolo26s.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:c8288f6a1076c19fd9c71c77b9abf62f2476ab85de3f310873e8069d197af878
-size 40261496
+oid sha256:fe561c64c597969297d3dc3ea9f9871957a9c42b4e67d3ed3a35bafb0d6130b6
+size 20165444

weights/yolo26x.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:227bd56f467af5c6f26b6064f513304e44cd7e694f4442da60c7d3e1222fde5f
-size 236507584
+oid sha256:78f4d0d310cd83830cb158b808196e2855937e08b495c4a00d472b0408ec074a
+size 118307852