folly::to<int_32t>(double) throws loss of precision exceptions

Here is a brief, reproducible code fragment example, written using `gtest` and `glog` syntax. It passes.

```
TEST_F(Test, TestCrashFollyConv) {
double y = 1.1;
int32_t x = (int32_t)y;
ASSERT_TRUE(x == 1);
x = floor(y);
ASSERT_TRUE(x == 1);
bool folly_threw = false;
try {
x = folly::to(y);
} catch (std::exception& e) {
LOG(ERROR) << "CAUGHT ERROR IN folly::to(double): " << e.what();
folly_threw = true;
}
ASSERT_TRUE(folly_threw);
}
```

Informally experimenting it appears an arbitrary number of `double` values fail to cast, raising this same error.

Of course there are multiple workarounds, such as `floor()` and `ceil()` in `cmath` specifically for this purpose. But the docs for `Conv` specifically claim "converts anything to anything" so I'm reporting it.

Taking a look at the code, float to int comparisons use lookup tables per digit that throw out of range exceptions if a position in the table is reached that doesn't match an expected position that would yield a digit. This is possibly the culprit, as under some failure scenarios I saw `std::range_error` raised, and in some saw the error message `to<>: loss of precision when converting 1.1 to type i`