Reduce code duplication without sacrificing performance

[WXbet]

Inspired by PR #4 (lukasz1992), this commit eliminates duplicated code across the stream cipher kernel and key scheduling functions while preserving the original runtime performance characteristics.

Stream cipher kernel (dvbcsa_bs_stream_kernel.inc):

• Replace the #ifdef DVBCSA_BS_STREAM_KERNEL_INIT / #else / #endif double-include trick with a single unified function _dvbcsa_bs_stream_cipher_kernel(const int init, regs).
• The function is marked static DVBCSA_INLINE inline, which maps to attribute((always_inline)) on GCC. Combined with the macro wrappers that pass literal constants (0 or 1), the compiler is guaranteed to inline and eliminate the dead branches via constant propagation — producing identical machine code to the original two-function approach.
• The sbox0 computation (4x unrolled Z/E carry chain) is intentionally kept unrolled. Extracting it into a loop function would break the regs->r "ultimate carry" calculation, which depends on tmp1/tmp3 values from the final iteration remaining in scope.

Stream cipher includes (dvbcsa_bs_stream.c):

• The .inc file is now included only once instead of the previous define/undef/include-twice pattern.

Kernel dispatch macros (dvbcsa_bs_stream_kernel.h):

• dvbcsa_bs_stream_cipher_kernel_init(regs) and dvbcsa_bs_stream_cipher_kernel(regs) are now macros that call _dvbcsa_bs_stream_cipher_kernel with compile-time constant 1 or 0.

Key scheduling (dvbcsa_key.c):

• dvbcsa_key_schedule_block() now delegates to dvbcsa_key_schedule_block_ecm(0, cw, kk), removing the duplicated loop body.

Little-endian load (dvbcsa_pv.h):

• dvbcsa_load_le64(p) is now a macro expanding to dvbcsa_load_le64_ecm(0, p). The compiler sees ecm=0 != 4 at compile time and eliminates the csa_block_perm_ecm lookup path entirely.