[Core] Fix a bug with fused conv2d (#2270)

BUG

- Fix a bug with activation being passed as null. (Happens from the layers side)
- Fix CI settings to reduce flakiness:
  - Lock Chrome to 67.0 on Windows to fix the WebGL problem on latest Chrome
  - Double the jasmine timeout from 5s to 10s to give us breathing room since sometimes tests in Windows take 5.1-5.3 secs

tfjs-core/karma.conf.js

@@ -152,7 +152,9 @@ module.exports = function(config) {
       win_10_chrome: {
         base: 'BrowserStack',
         browser: 'chrome',
-        browser_version: 'latest',
+        // Latest Chrome on Windows has WebGL problems:
+        // https://github.com/tensorflow/tfjs/issues/2272
+        browser_version: '77.0',
         os: 'Windows',
         os_version: '10'
       },

tfjs-core/src/jasmine_util.ts

@@ -14,11 +14,15 @@
  * limitations under the License.
  * =============================================================================
  */
+
+// tslint:disable-next-line:no-require-imports
+const jasmine = require('jasmine');
 import {KernelBackend} from './backends/backend';
 import {ENGINE} from './engine';
 import {env, Environment, Flags} from './environment';
 
 Error.stackTraceLimit = Infinity;
+jasmine.DEFAULT_TIMEOUT_INTERVAL = 10000;
 
 export type Constraints = {
   flags?: Flags,

tfjs-core/src/ops/fused_ops.ts

@@ -328,6 +328,7 @@ function conv2d_<T extends Tensor3D|Tensor4D>({
   activation?: Activation,
   preluActivationWeights?: Tensor
 }): T {
+  activation = activation || 'linear';
   if (shouldFuse(ENGINE.state.gradientDepth, activation) === false) {
     let result = unfusedConv2d(
         x, filter, strides, pad, dataFormat, dilations, dimRoundingMode);