Decompositions for reduceLogSum, reduceLogSumExp, and reduceSumSquare (#637)

* Decompositions for reduceLogSum, reduceLogSumExp, and reduceSumSquare

As noted in #209, not all frameworks/platform APIs support all reduce
operations, but these are easily emulated.

Decompositions are provided using the standard phrasing and style
currently used throughout the spec. In the future, we may alter the
wording to explicitly state that these decompositions can be used as
templates for implementations to follow when their underlying
frameworks/platform APIs do not support the operations.

Fixes #209

* Update index.bs

Co-authored-by: Ningxin Hu <[email protected]>

---------

Co-authored-by: Ningxin Hu <[email protected]>

Author: inexorabletash

index.bs

@@ -4929,6 +4929,27 @@ partial interface MLGraphBuilder {
     </div>
 </details>
 
+<div class="note">
+  <details open>
+    <summary>
+    The behavior of several reduction operations can be generically emulated from the usage of
+    other operations as follows. However, user agents typically have a more
+    efficient implementation for them, therefore their usage is encouraged from the
+    performance standpoint.
+    </summary>
+    <pre highlight="js">
+    // reduceLogSum(input, options)
+    return builder.log(builder.reduceSum(input, options));
+
+    // reduceLogSumExp(input, options)
+    return builder.log(builder.reduceSum(builder.exp(input), options));
+
+    // reduceSumSquare(input, options)
+    return builder.reduceSum(builder.pow(input, 2), options);
+    </pre>
+  </details>
+</div>
+
 ### relu ### {#api-mlgraphbuilder-relu-method}
 Compute the <a href="https://en.wikipedia.org/wiki/Rectifier_(neural_networks)">rectified linear function</a> of the input tensor.