Fix segfault in JS_NewSymbol with NULL description and align with ES semantics

[engsr6982]

Currently, calling JS_NewSymbol(ctx, NULL, false) triggers a segmentation fault because JS_NewAtom unconditionally calls strlen() on the description.

This PR fixes the crash and aligns the C-API with ECMAScript semantics:

1. Local Symbols (is_global = false): Passing NULL now creates a description-less Symbol (equivalent to Symbol()), correctly utilizing the internal sentinel JS_NewSymbolInternal(ctx, NULL, ...).
2. Global Symbols (is_global = true): Passing NULL now mimics Symbol.for(). Per the ES specification, Symbol.for() coerces undefined to the string "undefined". We fallback the NULL pointer to "undefined", avoiding any artificial C-level TypeError exceptions.

This makes JS_NewSymbol completely safe to use with optional C strings.

[bnoordhuis]

Looks fine to me but can you add a couple of regression tests to api-test.c?

[engsr6982]

Done. Added regression tests for both cases to api-test.c.