fix: use JPEG SOI/EOI marker scanning instead of multipart parsing

Recording now scans the fetch response for JPEG start-of-image (0xFFD8)
and end-of-image (0xFFD9) markers instead of parsing multipart boundaries.
This is more robust because:
- No dependency on boundary format quirks
- No crossOrigin attribute needed (no canvas tainting issues)
- JPEG SOI/EOI markers are reliable frame delimiters per the JPEG spec

Also added cancelRecording() cleanup at start of each new recording.

dist/index.html

@@ -5,7 +5,7 @@
     <link rel="icon" href="/favicon.ico">
     <meta name="viewport" content="width=device-width, initial-scale=1.0">
     <title>MJPEG Stream Viewer</title>
-    <script type="module" crossorigin src="/assets/index-3GYuRNvV.js"></script>
+    <script type="module" crossorigin src="/assets/index-C5dzEsTl.js"></script>
     <link rel="stylesheet" crossorigin href="/assets/index-DuJPD2XK.css">
   </head>
   <body>