Replace memset with fill to work for non-trivial scalars.

For custom scalars, zero is not necessarily represented by
a zeroed-out memory block (e.g. gnu MPFR). We therefore
cannot rely on `memset` if we want to fill a matrix or tensor
with zeroes. Instead, we should rely on `fill`, which for trivial
types does end up getting converted to a `memset` under-the-hood
(at least with gcc/clang).

Requires adding a `fill(begin, end, v)` to `TensorDevice`.

Replaced all potentially bad instances of memset with fill.

Fixes #2245.

test/OffByOneScalar.h

@@ -0,0 +1,28 @@
+
+// A Scalar with internal representation T+1 so that zero is internally
+// represented by T(1). This is used to test memory fill.
+// 
+template<typename T>
+class OffByOneScalar {
+ public:
+  OffByOneScalar() : val_(1) {}
+  OffByOneScalar(const OffByOneScalar& other) {
+    *this = other;
+  }
+  OffByOneScalar& operator=(const OffByOneScalar& other) {
+    val_ = other.val_;
+    return *this;
+  }
+  
+  OffByOneScalar(T val) : val_(val + 1) {}
+  OffByOneScalar& operator=(T val) {
+    val_ = val + 1;
+  }
+  
+  operator T() const {
+    return val_ - 1;
+  }
+ 
+ private:
+  T val_;
+};