Error message for non-scalar Tensor case.

[saeta]

When taking the gradient of a function, the function must return a scalar.
Because Tensors do not encode their shape or rank, this is checked dynamically
at runtime. Previously, we did not provide an error message, resulting in the
following error messages:

Precondition failed: file /swift-base/tensorflow-swift-apis/Sources/TensorFlow/Core/DifferentialOperators.swift, line 68
Current stack trace:
0    libswiftCore.so                    0x00007f1629f45830 swift_reportError + 50
1    libswiftCore.so                    0x00007f1629fb45d0 _swift_stdlib_reportFatalErrorInFile + 115
2    libswiftCore.so                    0x00007f1629edca7e <unavailable> + 3734142
3    libswiftCore.so                    0x00007f1629edcbf7 <unavailable> + 3734519
4    libswiftCore.so                    0x00007f1629cab10d <unavailable> + 1433869
5    libswiftCore.so                    0x00007f1629eb1a88 <unavailable> + 3558024
6    libswiftCore.so                    0x00007f1629caa569 <unavailable> + 1430889
9    repl_swift                         0x0000000000400490 <unavailable> + 1168
11   libswiftCore.so                    0x00007f1629caa569 <unavailable> + 1430889
14   repl_swift                         0x0000000000400490 <unavailable> + 1168
16   libswiftCore.so                    0x00007f1629caa569 <unavailable> + 1430889
17   libswiftTensorFlow.so              0x00007f16272250b0 <unavailable> + 2543792
18   libswiftTensorFlow.so              0x00007f1627098280 checkOk(_:file:line:) + 467
19   libswiftTensorFlow.so              0x00007f162709f480 TFE_Op.evaluateUnsafe() + 506
20   libswiftTensorFlow.so              0x00007f162709fcf0 TFE_Op.execute<A>(_:) + 132
21   libswiftTensorFlow.so              0x00007f16270a8984 <unavailable> + 985476
22   libswiftTensorFlow.so              0x00007f162713dc20 static Raw.matMul<A>(_:_:transposeA:transposeB:) + 1221
23   libswiftTensorFlow.so              0x00007f1627293c00 matmul<A>(_:transposed:_:transposed:) + 1427
24   libswiftTensorFlow.so              0x00007f16272f6210 _vjpMatmul<A>(_:transposed:_:transposed:) + 201
25   libswiftTensorFlow.so              0x00007f16273564b4 <unavailable> + 3794100
26   libswiftTensorFlow.so              0x00007f162731f960 AD__$s10TensorFlow5DenseV14callAsFunctionyAA0A0VyxGAGF__vjp_src_0_wrt_0_1 + 680
31   repl_swift                         0x0000000000400490 <unavailable> + 1168
Current stack trace:
	frame #4: 0x00007f162b926120 $__lldb_expr57`main at <Cell 8>:1

This change provides the user with a bit more context and prints out the shape
their computation produced to help them debug what might have gone wrong.

[saeta]

Note: I'm open to suggestions for what to call the "forward pass", if you have thoughts.

[Shashi456]

I know it's too much, but could we suggest or point to an example by which the person could debug his code?
Like for more look here type of thing. If there's one thing I don't like about error messages generally is, they only tell us what's wrong not how we could correct it.

[rxwei]

Thanks!

[eaplatanios]

Thanks Brennan!

[saeta]

Thank you @eaplatanios and @rxwei for speedy reviews! :-)

[saeta]

Note, looks like this is TF-320 (https://bugs.swift.org/browse/TF-320) Marking as fixed.