Switch back to subprocessing for llama.cpp

This should resolve a number of memory leak and stability defects by allowing
us to isolate llama.cpp in a separate process and shutdown when idle, and
gracefully restart if it has problems.  This also serves as a first step to be
able to run multiple copies to support multiple models concurrently.

llm/status.go

@@ -0,0 +1,42 @@
+package llm
+
+import (
+	"bytes"
+	"os"
+)
+
+// StatusWriter is a writer that captures error messages from the llama runner process
+type StatusWriter struct {
+	LastErrMsg string
+	out        *os.File
+}
+
+func NewStatusWriter(out *os.File) *StatusWriter {
+	return &StatusWriter{
+		out: out,
+	}
+}
+
+// TODO - regex matching to detect errors like
+// libcublasLt.so.11: cannot open shared object file: No such file or directory
+
+var errorPrefixes = []string{
+	"error:",
+	"CUDA error",
+	"cudaMalloc failed",
+	"\"ERR\"",
+}
+
+func (w *StatusWriter) Write(b []byte) (int, error) {
+	var errMsg string
+	for _, prefix := range errorPrefixes {
+		if _, after, ok := bytes.Cut(b, []byte(prefix)); ok {
+			errMsg = prefix + string(bytes.TrimSpace(after))
+		}
+	}
+	if errMsg != "" {
+		w.LastErrMsg = errMsg
+	}
+
+	return w.out.Write(b)
+}