[libc] Support underscores in NaN char sequences

[michaelrj]

Summary

Other libc implementations support underscores in NaN(n-char-sequence)
strings. Us not supporting that is causing fuzz failures, so this patch
solves the problem.

libc/src/__support/str_to_float.h

if (tolower(src[index + 1]) == nan_string[1] &&
       index += 3;
       BitsType nan_mantissa = 0;
       // this handles the case of `NaN(n-character-sequence)`, where the
       // n-character-sequence is made of 0 or more letters and numbers in any
       // order.
       if (src[index] == '(') {
         size_t left_paren = index;
         ++index;
-        while (isalnum(src[index]))
+        // Apparently it's common for underscores to also be accepted. No idea
+        // why, but it's causing fuzz failures.
+        while (isalnum(src[index]) || src[index] == '_')
           ++index;
         if (src[index] == ')') {
           ++index;
           if (isdigit(src[left_paren + 1])) {
             // This is to prevent errors when BitsType is larger than 64 bits,
             // since strtointeger only supports up to 64 bits. This is actually
             // more than is required by the specification, which says for the
             // input type "NAN(n-char-sequence)" that "the meaning of

libc/test/src/stdlib/strtof_test.cpp

TEST_F(LlvmLibcStrToFTest, NaNWithParenthesesInvalidSequenceTests) {
   run_test("NaN(123+asdf)", 3, 0x7fc00000);
   run_test("NaN(123", 3, 0x7fc00000);
 }
 
 TEST_F(LlvmLibcStrToFTest, NaNWithParenthesesValidSequenceInvalidNumberTests) {
   run_test("NaN(1a)", 7, 0x7fc00000);
   run_test("NaN(asdf)", 9, 0x7fc00000);
   run_test("NaN(1A1)", 8, 0x7fc00000);
+  run_test("NaN(why_does_this_work)", 23, 0x7fc00000);
+  run_test(
+      "NaN(1234567890qwertyuiopasdfghjklzxcvbnmQWERTYUIOPASDFGHJKLZXCVBNM_)",
+      68, 0x7fc00000);
 }