feat: lower LogicalNot, AllClose and Phase blocks in the native Expr compiler - #192

Summary

Closes the misc Nx.Block cluster on #188. Three new lower_block clauses, no new opcodes — each routes through primitives the IR already lowers, bit-identical to the Evaluator.

• Nx.Block.LogicalNot — emits the existing :logical_not opcode directly with a trailing coerce, mirroring Emily.Backend.native_logical_not/2.
• Nx.Block.AllClose — composes five existing primitives end-to-end, matching Emily.Backend.native_all_close/4 op-for-op: cast both to the merged float type, compute abs(a - b) <= atol + rtol * abs(b), optionally OR with isnan(a) AND isnan(b) for equal_nan: true, then reduce over every axis via :all. The trailing coerce produces {:u, 8}.
• Nx.Block.Phase — falls through to the block's composed expansion (atan2(imag(t), real(t))) via the TopK-style block-parameter seeding (block-local :parameter nodes are fresh and would otherwise resolve to the outer function's input slots). All three primitives in the expansion are lowered as of feat: lower 19 more unary ops in the native Expr compiler #189 / feat: lower atan2, quotient and logical_xor in the native Expr compiler #190, so the result is bit-identical to the Evaluator.

Probe (scripts/expr_op_coverage.exs) drops 12 → 9 misses; the misc Nx.Block section of #188 is now empty.

Test plan

• mix precommit clean (718 tests, 79 properties, 40 doctests, 0 failures — 6 new equivalence tests).
• Op-coverage probe shows all three moved from MISS → OK.
• New equivalence tests in compiler_equivalence_test.exs:
  • logical_not on both integer (s32) and float (f32) inputs, asserting the {:u, 8} predicate dtype.
  • all_close on close inputs (default tolerance → 1), far inputs (→ 0), custom rtol/atol inputs covering boundary cases (atol = 0.2 → close, atol = 0.05 → far), and the equal_nan: true branch (NaN compares equal to NaN, false is the default and returns 0).
  • phase across the four complex quadrants plus the real and imaginary axes (phase == 0 for positive real, ±pi for negative real, ±pi/2 for imaginary).
• No opcode changes — opcode parity stays at the current kOpcodeCount (115).

Follows the same pattern as #189 / #190 / #191.