refactor: remove backward-compat shim headers (core.h, einsum.h, core_py.h, einsum_py.h)

Deleted shims:
  numpy/core.h             → use numpy/numpy.h
  numpy/einsum.h           → use numpy/numpy.h
  pycpp/core_py.h          → use pycpp/pycpp.h
  pycpp/einsum_py.h        → use pycpp/pycpp.h

Updated stale references in bench/, example/, and detail headers.

Author: peng.li24

bench/bench_core.cpp

@@ -1,4 +1,4 @@
-#include "numpy/core.h"
+#include "numpy/numpy.h"
 #include "numpy/linalg.h"
 #include <benchmark/benchmark.h>
 #include <vector>

example/main.cpp

@@ -1,4 +1,4 @@
-#include <numpy/core.h>
+#include <numpy/numpy.h>
 #include <iostream>
 #include <vector>
 

numpy/core.h

@@ -2,8 +2,8 @@
 // ║  INTERNAL HEADER — DIRECT INCLUSION IS A COMPILE ERROR                 ║
 // ║                                                                          ║
 // ║  This file contains AVX-512 template specializations that override the  ║
-// ║  generic loops in core.h.  It is x86_64 + AVX-512F specific and must   ║
-// ║  be included INSIDE namespace numpy at the end of core.h — nowhere else.║
+// ║  generic loops in elementwise.h.  It is x86_64 + AVX-512F specific and  ║
+// ║  be included INSIDE namespace numpy at end of elementwise.h — nowhere else.║
 // ║                                                                          ║
 // ║  ✗  #include "numpy/detail/avx512_loops.h"     ← compile error                ║
 // ║  ✓  #include "numpy/numpy.h"             ← recommended entry point      ║

numpy/detail/avx512_loops.h

@@ -19,7 +19,7 @@
 //   np.linalg.norm(a) → sqrt(x.dot(x)) → same sdot_64_ / ddot_64_
 //
 // np.linalg.norm(a, axis=k) uses numpy's own pairwise sum — already
-// handled by norm_axis() in core.h, no BLAS needed.
+// handled by norm_axis() in linalg.h, no BLAS needed.
 //
 // The OpenBLAS library path is auto-discovered from /proc/self/maps
 // (numpy loads it when imported), so no compile-time link flag is needed.

numpy/detail/blas_bridge.h

@@ -6,14 +6,14 @@
 // ║  subject to change without notice.                                       ║
 // ║                                                                          ║
 // ║  ✗  #include "numpy/detail/npy_math_float.h"   ← compile error                ║
-// ║  ✓  #include "numpy/core.h"             ← only correct entry point      ║
+// ║  ✓  #include "numpy/numpy.h"            ← only correct entry point      ║
 // ╚══════════════════════════════════════════════════════════════════════════╝
 
 #pragma once
 
 #ifndef NUMPYCPP_INTERNAL_INCLUDE
 #  error "npy_math_float.h is an internal header — do not include directly. \
-Use #include \"numpy/core.h\" instead."
+Use #include \"numpy/numpy.h\" instead."
 #endif
 
 #include <cstdint>
@@ -24,7 +24,7 @@ namespace numpy {
 namespace detail {
 // Float32 bit-exact polynomial approximations — internal detail, do not use directly.
 // These replicate numpy's simd_exp_FLOAT, simd_log_FLOAT, simd_sincos_f32 algorithms.
-// Public API: use numpy::exp() etc. from core.h.
+// Public API: use numpy::exp() etc. from numpy.h.
 
 namespace {