Build multiple CPU variants and pick the best

This reduces the built-in linux version to not use any vector extensions
which enables the resulting builds to run under Rosetta on MacOS in
Docker.  Then at runtime it checks for the actual CPU vector
extensions and loads the best CPU library available

llm/ext_server.go

@@ -0,0 +1,17 @@
+//go:build !darwin
+
+package llm
+
+import (
+	"fmt"
+
+	"github.com/jmorganca/ollama/api"
+)
+
+func newDefaultExtServer(model string, adapters, projectors []string, opts api.Options) (extServer, error) {
+	// On windows and linux we always load the llama.cpp libraries dynamically to avoid startup DLL dependencies
+	// This ensures we can update the PATH at runtime to get everything loaded
+
+	// This should never happen as we'll always try to load one or more cpu dynamic libaries before hitting default
+	return nil, fmt.Errorf("no available default llm library")
+}