rlc: multi-user RLC kernel, bootrom toolchain fix, tile write-enable timing

[Aquaticfuller]

1. rlc: multi-user RLC kernel with use-case switching

Brings two things onto the multi-group kernel: the packet-assembly restructuring from PR #12
(@zouguoqi), and a multi-user extension on top of it.

Restructuring (PR #12): rlc_send_pkt / ue_status_rpt / pkt_production_and_recycle split
out of the consumer loop, per-packet TestDataStru, DFX counters, rate pacing.

Multi-user: one rlc_context_t per UE — rlc_ctx[NUM_USERS], sized from the data header's
ACTIVE_USER_NUMBER — with per-user list locks and per-user SN/ACK state. Producers route each
PDCP package to rlc_ctx[user_id] (the user_id was already in every descriptor and simply
ignored); consumers own a static subset of the entities. NUM_USERS == 1 reduces exactly to the
single-entity kernel, so TC1 behaviour is unchanged.

This covers TC2 (multi-user peak, 48 UEs, 800 B) from DP Introduction.docx; TC1 stays
selectable. The use case is chosen at build time by which data header is compiled
(M<users>_N<len>_K<pkgs> ⟷ data_<M>_<N>_<K>.h), so there is one kernel source, not a fork.

Fitted to this branch's core lists, not replacing them

main added explicit producer_core_ids[] / consumer_core_ids[] dispatch (#27), which is
strictly more general than the contiguous PRODUCER_CORE_NUM ranges the multi-user work was
originally written against. The lists are kept as the source of truth:

• a consumer's share of the entities follows its index in consumer_core_ids[], not
  core_id - PRODUCER_CORE_NUM, so producer/consumer cores need not be contiguous;
• the UE status task runs on producer_core_ids[0] instead of core 0 — with explicit lists
  core 0 need not be a producer, and gating on it would have dropped the ACK task entirely;
• producers_finished is removed as redundant: producer_done is already the count of producers
  that have finished (it additionally carries the this_core_done guard, so a slow producer
  cannot be abandoned mid-flight);
• PRODUCER_CORE_NUM / CONSUMER_CORE_NUM stay as fallbacks for headers generated before the
  lists existed, and as the pacing divisor.

Fixes carried along

• DATAHEADER now actually reaches rlc.c. It hard-coded #include "../data/data_1_1350_100.h",
  and since every generated header shares the PDCP_PKG_H guard, the per-variant define was a
  silent no-op — K10, K100 and K300 all compiled the same 100-packet dataset (byte-identical
  ELFs). The dead add_library(rlc ...) target is dropped with it: rlc.c needs the per-variant
  define, which a shared library target cannot provide (nothing linked it).
• PDU_STRIDE. The generator now pads the slot stride to 4 bytes. An 810 B PDU (800 + 10)
  leaves odd slots 2-byte aligned, which traps the word/vector payload copies; stride and copy
  length are now separate (descriptor pkg_length still carries 810).
• Rate pacing overflow. PRODUCER_CORE_NUM * PDU_SIZE * CPU_FREQENCY overflows int32
  (2·1360·1e9 wraps to 1285701632), so pacing was silently inert. Fixed with 64-bit math but left
  off by default (RLC_ENABLE_PACING=0 keeps the legacy expression verbatim) so existing
  numbers stay comparable — enabling it changes benchmark behaviour and should be a separate call.
• Optional payload integrity check behind RLC_SELF_CHECK.

2. Don't hardcode ETH toolchain for bootrom compilation (@jpf-h)

Johannes' fix from #25, carried here so it lands on main: the bootrom rule invoked
riscv -riscv64-gcc-9.5.0 …, which only exists on the ETH cluster, so make toolchain builds
could not build the bootrom. Now uses $(GCC_INSTALL_DIR)/bin/riscv32-unknown-elf-*.
Merged with this branch's dependency-list change to the same rule.

3. cachepool_tile: bank_we/bank_req as direct OR-reduce

Behaviour-preserving timing restructuring. bank_we/bank_req were produced by a priority
cascade plus a 4:1 sel_part_idx mux; they are value-identical to direct OR-reduces of the part
bits (the arbiter always selects a write-part when any write exists, else a read-part), so
computing them directly takes the write-enable off the critical-path tail
(data-bank we → meta-write clock-gate). addr/wdata/be still use the mux. Mirrors what
gen_unfolded_data_banks already did.