fix: preserve JSONSchemaExtend description; support {m,n} in tag patterns

[nghiack7]

Two related tag-parsing bugs, both in reflect.go:

---

Bug 1: JSONSchemaExtend description overwritten (#169)

structKeywordsFromTags unconditionally assigned jsonschema_description to t.Description, even when the tag was absent:

// before
t.Description = f.Tag.Get("jsonschema_description")  // "" when absent → erases prior description

When a type implements JSONSchemaExtend and sets a field's description, the property schema is first built by refOrReflectTypeToSchema (which calls JSONSchemaExtend), then structKeywordsFromTags runs and silently clears the description with an empty string.

Fix: only assign when the tag is non-empty:

if desc := f.Tag.Get("jsonschema_description"); desc != "" {
    t.Description = desc
}

---

Bug 2: Regex quantifiers {m,n} truncated in pattern tags (#181)

splitOnUnescapedCommas split on all unescaped commas, including those inside regex quantifiers like {0,2}:

// jsonschema:"pattern=^[A-Z]{0,3}$,maxLength=10"
// → pattern value truncated to "^[A-Z]{0"

The old backslash-escape convention (\,) still worked, but required users to manually escape every comma inside quantifiers, which was non-obvious and undocumented.

Fix: rewrite splitOnUnescapedCommas to track brace-nesting depth and skip commas inside {…}. The backslash-escape convention is still supported for backwards compatibility.

---

Tests added

• TestSplitOnUnescapedCommas: 3 new cases covering {m,n}, {1,{2,3}} nesting
• TestDescriptionPreservedFromJSONSchemaExtend: verifies description from JSONSchemaExtend survives structKeywordsFromTags
• TestPatternWithQuantifierComma: end-to-end schema generation with pattern=^[A-Z]{0,3}$

All existing tests pass.