ffi 2.16a green: parser + AST + sema for #jni_env(env) { body }

New `hash_jni_env` lexer token; `parsePrimary` dispatches to a small
`parseJniEnvBlock` that consumes `(env) { body }` and returns a new
`JniEnvBlock` AST node (env_expr + body block).

Sema's analyzeNode arm recurses into env + body inside a pushed
scope; findNodeAtOffset descends through both children for go-to-
definition.

Lowering treats it as a syntactic wrapper around the block: env is
evaluated for side effects, body lowers as a normal block. The TL
push/pop semantics (synthesizing the env stack so `#jni_call`'s env
arg can become optional) land in 2.16b.

`expectSemicolonAfter` recognises `jni_env_block` as block-form so
statement-position uses don't need a trailing `;` — matches `if` /
`while` / `for` / bare blocks.

Test runs through the block body and prints expected output; xfail
snapshot flips to green. 127/127 examples green.

src/lexer.zig

@@ -92,6 +92,7 @@ pub const Lexer = struct {
                 .{ "#extends", Tag.hash_extends },
                 .{ "#implements", Tag.hash_implements },
                 .{ "#jni_method_descriptor", Tag.hash_jni_method_descriptor },
+                .{ "#jni_env", Tag.hash_jni_env },
             };
             inline for (directives) |d| {
                 const keyword = d[0];