[SPARK-58443][SQL] Fix wrong results/exception in instr and substring_index for an Integer.MIN_VALUE position or count

[jiwen624]

What changes were proposed in this pull request?

Four backward-search paths negate a user-supplied int and use the result as a number of characters or delimiters to skip. -Integer.MIN_VALUE wraps back to Integer.MIN_VALUE, so the value stays negative: the search loop never runs and the out-of-range check that should have fired is bypassed.

Each site now performs the negation in long (-(long) start / -(long) count), so Integer.MIN_VALUE stays positive after negation and flows through the same range checks and search loops as every other negative value. No value is special-cased, and the loops still terminate as soon as the string runs out of characters or delimiters.

Site	Function	Collations	Behavior before
UTF8String.indexOf(pattern, start, occurrence)	instr	UTF8_BINARY, UTF8_LCASE (all-ASCII fast path)	behaves as if start = -1 instead of returning 0
UTF8String.subStringIndex	substring_index	UTF8_BINARY	NegativeArraySizeException
CollationAwareUTF8String.lowercaseSubStringIndex	substring_index	UTF8_LCASE	empty string
CollationAwareUTF8String.subStringIndex	substring_index	ICU	AssertionError under -ea; correct otherwise

The last row is latent: assertions are enabled in Spark's test JVMs but not in deployments, so it is hardening rather than a user-visible defect.

Why are the changes needed?

SELECT instr('abcabc', 'abc', -2147483648, 1);      -- returns 4, expected 0
SELECT substring_index('a.b.c', '.', -2147483648);  -- NegativeArraySizeException, expected a.b.c
SELECT substring_index(collate('a.b.c', 'UTF8_LCASE'), '.', -2147483648);  -- returns '', expected a.b.c

Does this PR introduce any user-facing change?

Yes - bug fix

How was this patch tested?

New test cases added

Was this patch authored or co-authored using generative AI tooling?

Yes

[uros-b]

Thank you @jiwen624!

[dongjoon-hyun]

+1, LGTM.

[HyukjinKwon]

0 blocking, 0 non-blocking, 0 nits.
Correct, minimal integer-overflow edge fix applied consistently across the binary and ICU-collation paths.

Verification

Confirmed the overflow root cause (int negation of MIN_VALUE) and that widening to long before negation removes it, making MIN_VALUE behave like any other large-magnitude negative per the function contract. Applied identically in both string implementations; tests cover MIN_VALUE across UTF8_BINARY/LCASE/UNICODE/UNICODE_CI.